JP2010016580A - Image processing apparatus and image processing method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an image processing apparatus that makes block boundaries inconspicuous in a processed image even in the case of dividing one screen into a plurality of blocks and performing processing in each block unit. <P>SOLUTION: The image processing apparatus is provided with: an image processing means 1000 for performing predetermined image processing of an image in a block unit obtained by dividing the image for one screen into a plurality of blocks and outputting processed image information in a unit for one screen; and a block dividing means 1005 for dividing the image for one screen into a plurality of blocks, generating an image of a block unit, transferring the generated image information of a block unit to the image processing means 1000, and also changing the division position of blocks on one screen to a different position in each screen. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to an image processing apparatus and method for processing, for example, an image such as a captured image from an image sensor in units of blocks obtained by dividing an image for one screen into a plurality of blocks.

  A block matching method for obtaining a motion vector between two screens from image information itself is an old technology.

  This block matching method is a method of calculating a motion vector between two screens between a reference screen that is a screen of interest and an original screen (referred to as a target screen) that is a source of motion of the reference screen. In the block matching method, for a block (target block and reference block) of a rectangular area having a predetermined size, the correlation between the reference screen and the target screen is calculated.

  There are both a case where the target screen is temporally preceding the reference screen and a case where the reference screen is temporally prior to the target screen. As an example of the former, for example, there is a case of motion detection in MPEG (Moving Picture Experts Group). Further, as the latter example, for example, there is a case of noise reduction by overlapping image frames described later.

  In this specification, the screen means an image composed of image data of one frame or one field. However, for the convenience of the following description in this specification, the screen is assumed to be composed of one frame. It will be called a frame. Therefore, the reference screen may be referred to as a reference frame, and the target screen may be referred to as a target frame.

  The motion vector detection method based on block matching has been developed mainly for pan / tilt detection of a television camera, subject tracking, and MPEG video coding. Since the beginning of the 90's, it has been applied to sensorless camera shake correction by superimposing images and noise removal during low-illuminance photography (Noise Reduction: hereinafter referred to as NR). -336121)).

  An NR processing apparatus using image superposition is, for example, as shown in FIG. 85, and includes a block division processing unit 1001 and an image processing unit 1000. Since this example is NR processing, the image processing unit 1000 includes a motion vector detection unit 1002, a motion compensation image generation unit 1003, and an image addition unit 1004.

  The image information TG of the target frame and the image information REF of the reference frame are supplied to the block division processing unit 1001, and for example, as shown in FIG. 86, each screen is divided into n pixels in the horizontal direction and in the vertical direction. Is divided every m lines. As a result, the image for one screen is divided into a plurality of blocks BK each consisting of horizontal direction × vertical direction = n pixels × m lines.

  Then, the block division processing unit 1001 sequentially supplies a block of the target frame (hereinafter referred to as target block) and a block of the reference frame (hereinafter referred to as reference block) to the motion vector detection unit 1002 in units of blocks. The

  The motion vector detection unit 1002 performs block matching between the target block and the reference block, and detects a motion vector MV for each block.

  The block-based motion vector MV detected by the motion vector detection unit 1002 is supplied to the motion compensation image generation unit 1003. The reference block is also supplied from the motion vector detection unit 1002 to the motion compensation image generation unit 1003.

  The motion compensated image generation unit 1003 generates a motion compensated image block in which the reference block is motion compensated using the motion vector. The motion compensation image is an image from which motion is removed using a motion vector.

  Then, the image information of the target block from the block division processing unit 1001 and the image information MVV of the motion compensation image block from the motion compensation image generation unit 1003 are supplied to the image addition unit 1004. The image adding unit 1004 superimposes the image information of the target block and the motion compensated image block from the motion compensated image generating unit 1003. The added image as a result of superposition is output from the image adding unit 1004 and stored in, for example, an image memory or recorded on a recording medium.

  The above processing is repeatedly executed for all blocks of one screen. When the processing for all the blocks of one screen is completed, the block division processing unit 1001 is supplied with the new target frame image information TG and the reference frame image information REF, and the above-described processing is repeated. It is.

  The superimposed image is more emphasized because the image itself should be the same image by motion compensation. On the other hand, since noise is random, it is canceled out by superposition. As a result, the superimposed image is reduced in noise.

The above-mentioned patent documents are as follows.
JP 2007-336121 A

  However, since the images are overlapped in units of blocks obtained by dividing one screen, the boundaries between the blocks may be noticeable. In particular, in a case where an output image is obtained by superimposing images of a plurality of frames of three or more and performing NR processing, an image obtained as a result of superimposing images of two frames is used as a target frame, and the next frame is There is a problem when images are superimposed as a reference image.

  That is, when the process of superimposing the first two images is process 1, and then the superimposed image is the target image and the process of superimposing the third and subsequent frames as the reference image is process 2, each time process 2 is repeated, Block boundaries may be noticeable.

  For example, if the block boundary Bd remains thin as shown in FIG. 87A in the image obtained as a result of superposition in the process 1, when the process 2 is executed, the block boundary Bd overlaps at the same position. As shown in FIG. 87 (B), it gradually remains dark.

  Such a block boundary is not limited to superimposition of images, but may occur when the image processing unit 1000 divides one screen into a plurality of blocks and processes the blocks.

  In view of the above points, the present invention provides an image processing apparatus and a block which makes block boundaries inconspicuous in a processed image even when one screen is divided into a plurality of blocks and processed in units of blocks. It aims to provide a method.

In order to solve the above problems, in the invention of claim 1,
Image processing means for performing predetermined image processing on the image in units of blocks obtained by dividing an image for one screen into a plurality of images, and outputting post-processing image information for one screen;
The image for one screen is divided into a plurality of blocks, the block unit image is generated, and the generated block unit image information is passed to the image processing means, and the block on the one screen is generated. There is provided an image processing apparatus comprising: a block dividing unit that changes the divided position to a different position for each screen.

  In the first aspect of the present invention, the block division position on one screen is changed to a different position for each screen, so that the block division position becomes inconspicuous.

  According to the present invention, since the block division position on one screen is changed to a different position for each screen, the block division position becomes inconspicuous.

  Embodiments of an image processing apparatus and an image processing method according to the present invention will be described below with reference to the drawings.

[First Embodiment]
FIG. 1 is a block diagram showing a schematic configuration example of an embodiment when an image processing apparatus according to the present invention is applied to the NR processing described in the background art. In FIG. 1, the same parts as those in FIG. 85 described above are denoted by the same reference numerals.

  In this embodiment, a division position shift control unit 1005 is connected to the block division processing unit 1001. The division position shift control unit 1005 is supplied with a timing signal FLv having a frame period from the timing signal generation unit 1006. Of course, the timing signal FLv of this frame period is synchronized with the target frame and the reference frame.

  Note that the control unit (not shown) that controls the operations of the motion vector detection unit 1002, the motion compensation image generation unit 1003, and the image addition unit 1004 of the image processing unit 1000 also uses the timing signal FLv from the timing signal generation unit 1006. A synchronized timing signal is provided. Each unit of the image processing unit 1000 is controlled by a control signal synchronized with the timing signal from the control unit to execute processing.

  The division position shift control unit 1005 receives the timing signal FLv of the frame period from the timing signal generation unit 1006, and controls the division processing in the block division processing unit 1001 so as to change the division block position for each frame.

  In the block division processing unit 1001, for the received target frame and reference frame image information, first, as shown in FIG. 2A, for example, the upper left corner is set as the division start point Po, and every n pixels in the horizontal direction. Divide and divide every m lines in the vertical direction.

  The image information for each block Bk generated by the division processing in the block division processing unit 1001 is sequentially supplied to the image processing unit 1000. In this example, the image processing unit 1000 sequentially performs motion vector detection processing, motion compensation image generation processing, and image addition processing in units of block Bk, as in the case of FIG.

  When the above-described image processing for all the blocks Bk for one frame is completed in the image processing unit 1000, the block division processing unit 1001 sets the starting point for image division based on the control instruction of the division position shift control unit 1005. The start point Po is changed to the start point Pa. Then, in the block division processing unit 1001, for the received image information of the second target frame and reference frame, as shown in FIG. 2B, starting from the division start point Pa, every n pixels in the horizontal direction. And is divided every m lines in the vertical direction. That is, in this embodiment, the size of the block Bk is not changed.

  At this time, as shown with a shadow line in FIG. 2B, for a block including a region that protrudes from the image frame of the original one-frame image region, an image region excluding a region portion that protrudes from the image frame. Only the image processing target is changed, and the block size is changed.

  Alternatively, a block including a region that is generated by interpolating pixel data of a region that protrudes from the image frame of the original one-frame image region and that protrudes from the image frame of the original one-frame image region is also a block for image processing. The size of may not be changed.

  Furthermore, a block including an area that protrudes from the image frame of the original image area of one frame may be excluded from the image processing target.

  Thus, the image information of the second target frame and reference frame is divided into blocks starting from a division start point Pa different from the first division start point Po, and the image processing unit 1000 performs the above-described division on the divided blocks. Image processing is performed.

  The image information of the third and subsequent target frames and reference frames is also divided into blocks without changing the block size, starting from different division start points. Then, the image processing unit 1000 performs the above-described image processing on the divided blocks.

  In addition, the change of the position of the division start point of block division is performed within one block.

  As described above, in this embodiment, since the block division position of the screen is changed for each frame, the boundary between the blocks can be made inconspicuous in the superimposed image.

  In particular, if the process of superimposing the first two images is process 1, then the superimposed image is the target image, and the process of superimposing the third and subsequent frames as the reference image is process 2, the process 2 may be repeated. , Block boundaries can be prevented from being noticeable.

  That is, in the conventional case, since the block division position is fixed, when the block boundary Bd remains thin in the image obtained as a result of the superimposition in the processing 1, the block boundary as shown in FIG. Will stand out.

  On the other hand, according to this embodiment, even when the block boundary Bdo remains thin as shown in FIG. 3A in the image obtained as a result of the process 1, when the process 2 is executed, the image becomes thin. A block boundary Bda that remains lightly appears at a position different from the remaining block boundary Bdo. For this reason, the block boundary Bdo and the block boundary Bda do not overlap, and the block boundary becomes inconspicuous.

[Example of block division position change control]
Next, some examples of block division position change control will be described. The examples described below are examples in which the block division position is changed and controlled by changing the position of the division start point (division start point) of block division.

<First example>
In the first example, as shown by an arrow AR in FIG. 4A, the division start point of block division is moved from the first division start point Po at the upper left corner of the frame image in the diagonal direction of the block Bk. To do. In this first example, the block Bk has a size of horizontal direction × vertical direction = 8 pixels × 8 lines, and every time image processing for one screen is completed, one block in the horizontal direction and one in the vertical direction. The division start point is moved by one line.

  That is, for the first target frame, block division is performed with 8 pixels in the horizontal direction and 8 lines in the vertical direction from the division start point Po. When the image processing for one frame in the block unit is completed, for the next target frame, as shown in FIG. 4B, the division start point Po is shifted by one pixel in the horizontal direction and one line in the vertical direction. The division start point P1 is set. Then, block division is performed with 8 pixels in the horizontal direction and 8 lines in the vertical direction starting from the division start point P1.

  For the next target frame, as shown in FIG. 4C, the division start point P1 is further set to one pixel in the horizontal direction and one line in the vertical direction. Therefore, two pixels in the horizontal direction from the division start point Po. The division start point P2 that is shifted by two lines in the vertical direction is set.

  Thereafter, every time image processing for one screen is completed, the division start point is moved by one pixel in the horizontal direction and one line in the vertical direction. That is, assuming that the division start point of block division is the pixel coordinate of the left corner position of the frame image (0, 0), the transition of the division start point in pixel units is Po → P1 → P2 → P3... → P7 = (0 , 0) → (1,1) → (2,2) → (3,3)... → (7,7).

  As shown in FIG. 4D, the coordinate start point of the division start point P7 of the seventh frame corresponds to the vertex position of the diagonal position of the block Bk. Therefore, in the next frame, as shown by a dotted arrow in FIG. 4A, the first division start point Po is returned. Hereinafter, FIGS. 4A to 4D are repeated.

  Of course, instead of shifting one pixel and one line for each frame, it is possible to shift a plurality of pixels and a plurality of lines for each frame. For example, when shifting by 2 pixels × 2 pixels, the transition of the division start point is Po → P2 → P4 → P6 → P1 → P3 → P5 → P7 = (0, 0) → (2, 2) → (4 4) → (6,6) → (1,1) → (3,3) → (5,5) →→ (7,7).

<Second example>
This example is a modification of the first example. In the first example, the processing of image information for one frame is completed for each pixel in the horizontal direction and one line in the vertical direction, and the division start point is shifted when processing the next frame. In the second example, the division start point is shifted in consideration of the golden ratio instead of one pixel in the horizontal direction and one line in the vertical direction.

Here, the golden ratio is a ratio that is mathematically most beautiful, and is a numerical value of (5 ^1/2 ± 1) /2≈1.618 or 0.618. It is well known that this ratio (1: 1.618 or 1: 0.618) is often found in the natural world (how to attach petals, the number of sunflower seeds, etc.) and ancient objects (Parthenon and Arc de Triomphe). It has been.

  One of the reasons why the numerical value expressed by the golden ratio is mathematically beautiful is that the phase from “0” to “1” is added in increments of 0.618 by the golden ratio. There is a feature that is. This can also be understood from the fact that when a certain length is continuously marked with the golden ratio, it can be marked without any bias.

  A second example is a method of sequentially shifting the division start point using the feature of the golden ratio. An example is shown in FIG. In the example of FIG. 5, one block Bk is a rectangle of 8 pixels × 8 lines.

  In the example of FIG. 5, as in the first example, the division start point is shifted frame by frame in the diagonal direction of one block indicated by the arrow AR. In this case, in this second example, as shown in FIGS. 5 (A), (B), (C), (D), coordinates (7, 7) from coordinates (0, 0) on the diagonal are shown. Until the first division start point Po (coordinate is (0, 0)), the division start point is sequentially changed by the golden ratio.

  In this case, the length of the diagonal line of one rectangular block is 8√2 from the three-square theorem. The golden ratio times the length of the diagonal line is (8√2 × (√5-1) / 2) ≈7. And, by replacing it with coordinates from the three-square theorem and similar forms, (4.9, 4, 9) ≒ (5, 5), and the diagonal coordinates (0, 0) to (7, 7) In the meantime, it is shifted by 5 pixels in the horizontal direction and 5 lines in the vertical direction.

  As described above, when the division start point Po of the block division is the left corner position of the frame image and the pixel coordinates are (0, 0), the transition of the division start point in pixel units is, for example, Po → P1 → P2 → P3 → P4 → P5 → P6 → P7 → P0 →... = (0,0) → (5,5) → (2,2) → (7,7) → (4,4) → (1,1) → (6, 6) → (3, 3) → (0, 0).

  Thus, in the example of FIG. 5, since the division start point is shifted in units of pixels, the division start point is repeatedly changed to be located at the different eight point positions as described above.

<Third example>
The first example and the second example are examples in which the division start point is gradually shifted in the diagonal direction of the rectangular block Bk, but the direction in which the division start point is shifted is not limited to this.

  The third example shows another example of the direction in which the division start point is shifted. In this example, as shown in FIGS. 6A and 6B, the division start point is set in the upper left corner in one block Bk. From the starting point Po, the rectangular spiral shape is changed. When the end point of the division start point to be changed into a spiral shape is Pe, the division start point Pe is returned to the division start point Po after the end point of the division start point Pe.

  Also in the third example, the division start point for each frame may be changed for each pixel, for each line, for each of a plurality of pixels, and for each of a plurality of lines, and the position of the division start point is changed according to the golden ratio. May be.

<Fourth example>
This fourth example is another example of the shifting direction. In this example, as shown in FIGS. 7A and 7B, the division start point is changed from the starting point Po at the upper left corner to a zigzag shape within one block Bk. Also in this fourth example, when the end point of the division start point to be changed to the zigzag shape is Pe, the division start point Pe is returned to the division start point Po after the end point of the division start point Pe.

  Also in the fourth example, the division start point for each frame may be changed for each pixel and for each line, for each of a plurality of pixels and for each of a plurality of lines, and the position of the division start point is changed according to the golden ratio. May be.

<Fifth example>
This fourth example is another example of the shifting direction. In this example, as shown in FIGS. 8A and 8B, the division start point is shifted in the line direction from the starting point Po at the upper left corner in one block Bk. In a plurality of lines, the lines are shifted in the line direction. Then, after the division start point has been shifted by one line in the line direction, the changing process of shifting the line in the line direction is repeated for every other line.

  Also in this fifth example, when the end point of the division start point to be changed is Pe, the division start point Po is returned to the division start point Po next to the end point of the division start point Pe.

  Also in the fifth example, the division start point for each frame may be changed for each pixel or a plurality of pixels, or the position of the division start point may be changed based on the golden ratio.

[Second Embodiment]
Next, a more specific embodiment when the image processing apparatus and the image processing method according to the present invention are applied to an imaging apparatus will be described as a second embodiment with reference to the drawings.

  The image pickup apparatus according to the second embodiment performs noise reduction of an image by superimposing a plurality of still images picked up by an image pickup element (imager) while aligning them.

  The environment in which the imaging apparatus is held by hand for shooting is an environment in which camera shake is likely to occur. At the time of handheld shooting of such an imaging apparatus, for example, when the still image is continuously shot, the first frame is set as the target frame, and the second frame and thereafter are set as reference frames.

  If the reference frame can be aligned with the target frame by performing deformation processing with translation, rotation, and scaling on the entire reference frame, the deformation processing applied to the entire reference frame is global. It is motion. The global motion usually indicates the movement and amount of movement of a background still image (hereinafter abbreviated as background) of the image.

  Camera shake means that a global motion has occurred with respect to the target frame by the amount that the reference frame has undergone image deformation accompanied by translation, rotation, and enlargement / reduction with respect to the target frame.

  In this embodiment, for the sake of simplicity of explanation, the first photographed still image is used as a reference (target frame). But this is not essential. The global motion may be redefined with an arbitrary nth frame as a target frame and an arbitrary m (n ≠ m) th frame as a reference frame. Further, global motion may be defined between two frame images that are temporally different in a moving image. Further, the global motion processing may be applied to only a part of the image of one frame without targeting all the effective image frames in one frame of the captured image.

  When global motion is obtained, a motion vector corresponding to the motion applied to the entire screen between the target frame and the reference frame for each coordinate position on the screen can be calculated from the obtained global motion. . A motion vector corresponding to the motion applied to the entire screen is hereinafter referred to as a global motion vector. This global motion vector can be obtained for each of a large number of target blocks set on the screen. A global motion vector can also be obtained for each pixel.

  And if a global motion vector can be calculated | required correctly based on global motion, camera shake can be correct | amended favorably. Alternatively, the frame images can be overlaid while correcting the amount of camera shake in the global motion vector.

  In the following description, NR (Noise Reduction) is to reduce noise by superimposing a plurality of images using motion vector detection and motion compensation (aligning two frames of images with motion vectors). Call. An image whose noise is reduced by NR is referred to as an NR image.

  In this embodiment, one screen is divided into a large number of blocks (target blocks to be described later), and the motion vector for each block is detected by the block matching method described above. The motion vector detected for each block by the block matching method is referred to as a local motion vector in this specification.

  In this embodiment, the reliability index of the local motion vector is detected as described later using the correlation value in the block matching method. In this embodiment, the global motion is calculated from only the highly reliable local motion vectors detected, and the global motion vector for each block is detected from the calculated global motion.

[Outline of block matching]
9 to 14 are diagrams for explaining the outline of the block matching method. In the block matching method described here, for example, as shown in FIG. 9A, a block of a rectangular area of a predetermined size, which includes a plurality of pixels in the horizontal direction and a plurality of lines in the vertical direction, on the target frame 100, respectively. (Target block) 102 is assumed. Also, a block (reference block) having the same size as the target block is set in the reference frame 101.

  In block matching, a reference block having a high correlation with the target block 102 is searched from the reference frame 101. As a result of this search, the reference block 103 (see FIG. 9B) detected in the reference frame 101 as having the highest correlation is referred to as a motion compensation block. Further, the amount of displacement between the target block 102 and the motion compensation block 103 (reference block having the highest correlation) is referred to as a motion vector (see reference numeral 104 in FIG. 9B).

  The motion vector 104 corresponding to the positional deviation (including the positional deviation amount and the positional deviation direction) between the target block 102 and the motion compensation block 103 is the position of the projected image block 109 of the target block in the reference frame 101 and the motion compensation. This corresponds to a positional deviation from the position of the block 103. Here, as the position of the projection image block 109 and the position of the motion compensation block 103, for example, the center position of the block is used.

  The projected image block 109 of the target block 102 is assumed at the same position as the position of each target block 102 of the target frame 100 in the reference frame 101.

  An outline of the block matching process will be described.

  First, as shown by a dotted line in FIG. 10, a projected image block 109 of the target block 102 is assumed in the reference frame 101 at the same position as the target block 102 of the target frame 100. The coordinates of the center of the projected image block 109 of the target block 102 are set as the motion detection origin 105. Then, assuming that the motion vector 104 exists within a certain range from the motion detection origin 105, a predetermined range centered on the motion detection origin 105 is set as the search range 106 (see the one-dot chain line in FIG. 10). .

  Next, a block (referred to as a reference block) 108 having the same size as the target block 102 is set on the reference screen. Then, the position of the reference block 108 is moved in the set search range 106 in units of one pixel or a plurality of pixels, for example, in the horizontal direction and the vertical direction. Accordingly, a plurality of reference blocks 108 are set in the search range 106.

  Here, the reference block 108 is moved within the search range 106. In this example, since the motion detection origin 105 is the center position of the target block, the center position of the reference block 108 is moved within the search range 106. It means to move. Therefore, the pixels constituting the reference block 108 may protrude beyond the search range 106.

  Then, for each reference block 108 to be set in the search range, a vector (referred to as a reference vector) 107 (refer to FIG. 10) representing the positional deviation amount and the positional deviation direction between each reference block 108 and the target block 102. Set. Then, the correlation between the image content of the reference block 108 at the position indicated by each reference vector 107 and the image content of the target block 102 is evaluated.

  As shown in FIG. 11, the reference vector 107 is a vector (Vx, Vy) where the horizontal displacement (X direction) of the reference block 108 is Vx and the vertical displacement (Y direction) is Vy. )It can be expressed as. When the position coordinates (for example, center position coordinates) of the reference block 108 and the position coordinates (for example, center position coordinates) of the target block 102 are the same, the reference vector 107 is represented as a vector (0, 0).

  For example, when the reference block 108 is at a position shifted by one pixel in the X direction from the position of the target block 102, the reference vector 107 is a vector (1, 0). As shown in FIG. 12, when the reference block 108 is located at a position shifted by 3 pixels in the X direction and 2 pixels in the Y direction from the position of the target block 102, the reference vector 107 is a vector (3, 2 )

  That is, as shown in the example of FIG. 12, the reference vector 107 means a positional deviation (a vector including a positional deviation amount and a positional deviation direction) between each corresponding reference block 108 and the target block 102. To do. In FIG. 12, the positions of the target block 102 and the reference block 108 are the center positions of the respective blocks.

  The reference block 108 moves within the search range 106. In this case, the center position of the reference block 108 moves within the search range 106. The reference block 108 includes a plurality of pixels in the horizontal direction and the vertical direction. Accordingly, the maximum range in which the target block 102 and the reference block 108 to be subjected to block matching processing move is a matching processing range 110 wider than the search range 106, as shown in FIG.

  Then, the position of the reference block 108 detected as having the strongest correlation with the image content of the target block 102 is detected as the position of the target block 102 of the target frame 100 in the reference frame 101 (position after moving). The detected reference block is set as the motion compensation block 103 described above. Then, the displacement amount between the detected position of the motion compensation block 103 and the position of the target block 102 is detected as a motion vector 104 as an amount including a direction component (see FIG. 9B). ).

  Here, the correlation value indicating the strength of correlation between the target block 102 and the reference block 108 moving in the search range 106 is basically calculated using the pixel values corresponding to the target block 102 and the reference block 108. Is done. As the calculation method, a method using a root mean square and various other methods have been proposed.

  Among them, as a correlation value generally used when calculating a motion vector, for example, the absolute value of the difference between the luminance value of each pixel in the target block 102 and the luminance value of each corresponding pixel in the reference block 106 is used. The sum of values for all pixels in the block is used (see FIG. 13). The sum of the absolute values of the differences is called the difference absolute value sum. Hereinafter, this sum of absolute differences is referred to as a SAD (Sum of Absolute Difference) value.

  When the SAD value is used as the correlation calculation result, the smaller the SAD value, the stronger the correlation. Therefore, among the reference blocks 108 that move in the search range 106, the reference block 108 at the position where the SAD value is the smallest is the strongest correlation reference block with the strongest correlation. This strongest correlation reference block is detected as the motion compensation block 103, and the amount of displacement of the detected motion compensation block 103 relative to the position of the target block 102 is detected as a motion vector.

  In general, the correlation is stronger when the correlation value is larger. When the SAD value is used as the correlation calculation result, it can be said that the smaller the SAD value, the larger the correlation value and the stronger the correlation.

  As described above, in block matching, the amount of positional deviation of each of the plurality of reference blocks 108 set in the search range 106 with respect to the position of the target block 102 is expressed by the reference vector 107 as an amount including a direction component. The The reference vector 107 of each reference block 108 has a value corresponding to the position of the reference block 108 on the reference frame 102. As described above, in block matching, the reference vector of the reference block 108 in which the SAD value that is the correlation value has the minimum value is detected as the motion vector 104.

  Therefore, in block matching, the SAD value between each of the plurality of reference blocks 108 set in the search range 106 and the target block 102 (hereinafter referred to as the SAD value for the reference block 108 for the sake of simplicity) First ask.

  Next, as shown in FIG. 14, the obtained SAD value is stored in the memory in correspondence with each reference vector 107 corresponding to the position of each reference block 108. The motion vector 104 is detected by detecting the reference block 108 having the smallest SAD value from the SAD values for all the reference blocks 108 stored in the memory. For the sake of simplicity, the reference vector 107 corresponding to the position of the reference block 106 is hereinafter referred to as the reference vector 107 of the reference block 108.

  Corresponding to each of the reference vectors 107 corresponding to the positions of the plurality of reference blocks 108 set in the search range 106, a correlation value (in this example, SAD value) for each reference block 108 is stored. This is called a correlation value table. In this example, since the SAD value which is the sum of absolute differences is used as the correlation value, this correlation value table will be referred to as a difference absolute value sum table (hereinafter referred to as SAD table).

  This is shown in the SAD table TBL of FIG. 14, and the correlation value (SAD value in this example) for each reference block 108 in this SAD table TBL is referred to as a correlation value table element. In the example of FIG. 14, the SAD value indicated by reference numeral 111 is the SAD value when the reference vector is the vector (0, 0). In the example of FIG. 14, since the minimum value of the SAD value is “7” when the reference vector is the vector (3, 2), the motion vector 104 to be obtained is (3, 2).

  In the above description, the positions of the target block 102 and the reference block 108 mean arbitrary specific positions of these blocks, for example, the center position. The reference vector 107 indicates a shift amount (including direction) between the position of the projection image block 109 of the target block 102 and the position of the reference block 108 in the reference frame 102.

  The reference vector 107 corresponding to each reference block 108 is a positional shift of each reference block 108 from the position of the projection image block 109 corresponding to the target block 102 on the reference frame 101. Therefore, when the position of the reference block 108 is specified, the value of the reference vector is also specified corresponding to the position. Thus, when the address of the correlation value table element of the reference block in the memory of the SAD table 110 is specified, the corresponding reference vector is specified.

  The SAD value may be calculated in parallel for two or more target blocks at the same time. As the number of target blocks to be processed simultaneously increases, the processing speed increases. However, since the hardware scale for calculating the SAD value increases, there is a trade-off between speeding up the processing and increasing the circuit scale.

[Block Matching of Embodiment]
The above description of block matching is a description of motion vector calculation for one target block. Since the ratio of the target block to the entire target frame is usually small, it is generally difficult to obtain a motion (global motion) for the target frame based on the one target block.

  In this embodiment, as shown in FIG. 15, the target frame 100 is divided into a plurality of target blocks 102 having a relatively small size, for example, 8 pixels × 8 lines. In this embodiment, the division position of the target block 102 is changed every time the block matching process for the image for one frame is completed as described in the first embodiment.

  In processing within one frame, first, a motion vector (local motion vector) 104B for each of a plurality of target blocks 102 obtained by block division is obtained. At this time, in this embodiment, an index indicating the reliability of each obtained local motion vector is calculated together.

  Then, from the local motion vector reliability index, only a highly reliable local motion vector is extracted from a plurality of local motion vectors obtained for the target frame.

  Next, global motion is calculated only from the extracted local motion vector with high reliability. Then, a global motion vector for each target block is calculated using the calculated global motion.

  Then, the calculated global motion vector and the local motion vector for each target block are compared, and based on the comparison result, it is evaluated whether each target block is a background portion or a moving subject portion. In this embodiment, the calculated global motion vector and the local motion vector for each target block are compared to determine the degree of coincidence between them. Then, as the determination result, an index value indicating the degree of coincidence between the local motion vector and the global motion vector for each target block is calculated. This index value is referred to as a hit rate in this specification.

  Such evaluation determination is performed in consideration of the influence of noise included in an image on a correlation value calculated by block matching.

  When the global motion vector and the local motion vector for the target block match, it can be determined that the target block is the background image portion. Accordingly, the index value of the degree of coincidence indicates the degree (background coincidence degree) of whether or not the target block image coincides with the background image portion.

  If the target block is a background image portion, it is a still image portion, so the motion compensated image can be added to the target block image at a rate of 100% (ratio of 1: 1), and NR ( Noise reduction effect can be maximized.

  If the degree of background matching is low, the target block is expected to be a moving subject part. Therefore, by reducing the addition ratio of the motion compensated image to the image according to the degree of background matching, multiple exposure is achieved. Can be reduced.

  Here, if the global motion vector and the local motion vector do not match, it may be determined that all of the target blocks are moving subject portions without considering image noise. In this case, the correlation value for the reference block corresponding to the local motion vector is maximum (the SAD value is minimum), and is larger than the correlation value for the reference block corresponding to the global motion vector (the SAD value is small).

  However, noise is generally included in an image such as a captured image. Considering this image noise, the target block may be the background portion even when the global motion vector and the local motion vector do not match. In such a target block, the difference between the correlation value for the reference block corresponding to the local motion vector and the correlation value for the reference block corresponding to the global motion vector is smaller than the image noise. Conceivable.

  Therefore, in this embodiment, the correlation value for the reference block corresponding to the global motion vector is corrected to a value reflecting the noise of the image, and the corrected correlation value and the reference block corresponding to the local motion vector are corrected. Compare with the correlation value for. When the corrected correlation value is larger (the SAD value is smaller), the target block is evaluated as a background image portion. That is, in this embodiment, the background matching degree is evaluated based on the corrected correlation value. At this time, it is considered that the global motion vector matches the original local motion vector for the target block.

  In this embodiment, when it is determined that the target block is a background image portion as a result of evaluating the degree of background matching, a global motion vector is output as an output motion vector for the target block. If it is determined that the target block does not match the background image portion as a result of the background matching degree evaluation, a local motion vector is output as an output motion vector for the target block. In this embodiment, the output motion vector is a subsequent processing motion vector, and in this embodiment, an NR processing motion vector.

  Needless to say, when the global motion vector and the local motion vector completely match, either the global motion vector or the local motion vector may be used as the output motion vector.

  In this embodiment, using the output motion vector in units of target blocks obtained as described above, the reference frame is aligned in units of blocks with respect to the target frame, and a motion compensated image (motion compensation frame) is obtained. Is generated. Then, the NR image is generated by superimposing the target frame and the motion compensation frame.

  In the imaging apparatus of this embodiment, in still image shooting, as shown in FIG. 16, a plurality of still images are captured at high speed, and the first still image captured image is used as the target frame 100. . Then, for the second and subsequent sheets, a predetermined number of still image captured images are used as a reference frame 101, and the superimposed images are recorded as still image captured images.

  That is, when the photographer depresses the shutter button of the image pickup apparatus, the predetermined number of still images are taken at high speed. Then, a plurality of still images (frames) taken later in time are superimposed and recorded on the first still image (frame) taken.

  Although not targeted in this embodiment, at the time of moving image shooting, as shown in FIG. 17, the current frame image output from the image sensor is set as the target frame 100 image, and the past image of the previous frame is referred to. The image of the frame 101 is assumed. That is, at the time of moving image shooting, the image of the previous frame of the current frame is superimposed on the current frame in order to reduce noise of the image of the current frame.

[Example of hardware configuration of imaging device]
FIG. 18 shows a block diagram of an example of an imaging apparatus as an embodiment of the image processing apparatus of the present invention.

  As shown in FIG. 18, in the imaging apparatus of this embodiment, a CPU (Central Processing Unit) 1 is connected to a system bus 2. Further, the system bus 2 is configured by connecting an imaging signal processing system 10, a user operation input unit 3, an image memory unit 4, a recording / reproducing device unit 5, and the like. In this specification, the CPU 1 includes a ROM (Read Only Memory) and a work area RAM (Random Access Memory) that store programs for performing various software processes, although not shown. .

  The CPU 1 is configured to control the operation of each processing unit through the system bus 2.

  Upon receiving an imaging recording start operation through the user operation input unit 3, the imaging signal processing system of the imaging apparatus in FIG. 18 performs recording processing of captured image data as described later. Further, in response to an operation for starting reproduction of the captured image through the user operation input unit 3, the imaging signal processing system 10 of the imaging apparatus in FIG. 18 reproduces the captured image data recorded on the recording medium of the recording / reproducing apparatus 5. Perform processing.

  As shown in FIG. 18, in the imaging signal processing system 10, incident light from a subject through a camera optical system (not shown) including an imaging lens 10 </ b> L is applied to the imaging element 11 and imaged. In this example, the imaging device 11 is configured by a CCD (Charge Coupled Device) imager. Note that the image sensor 11 may be a CMOS (Complementary Metal Oxide Semiconductor) imager.

  In the imaging apparatus of this example, when an imaging recording start operation is performed, an image input through the lens 10L is converted into a captured image signal by the imaging element 11. The signal synchronized with the timing signal from the timing signal generator 12 is a RAW signal (raw signal) in a Bayer array composed of three primary colors of red (R), green (G), and blue (B). An analog image signal is output from the image sensor 11. The output analog imaging signal is supplied to the preprocessing unit 13, subjected to preprocessing such as defect correction and γ correction, and supplied to the data conversion unit 14.

  The data converter 14 converts the analog imaging signal, which is a RAW signal input thereto, into a digital imaging signal (YC data) composed of a luminance signal component Y and a color difference signal component Cb / Cr. A digital imaging signal from the data conversion unit 14 is written in the image memory unit 4 based on a shooting instruction through the user operation input unit 3.

  In other words, if the shooting instruction through the user operation input unit 3 is a still image shooting instruction by pressing the shutter button, the digital imaging signal from the data conversion unit 14 is written in the frame memories 41 to 4N. In this case, the digital image pickup signal from the data conversion unit 14 includes the first to Nth (N is the number of still images to be superimposed) frame memories 41 to 41 in the image memory unit 4 for the plurality of frames to be superimposed. Written to 4N.

  In this example, the image data of the first frame when the shutter button is pressed is written in the first frame memory 41 as the image data of the target frame. Then, the data of the second and subsequent frame images are sequentially written in the second to Nth frame memories 42 to 4N as reference frame image data, respectively.

  After images for a plurality of frames are written in the image memory unit 4, the image data of the target frame and the image data of the reference frame are read by the motion vector calculation unit 15.

  The motion vector calculation unit 15 functionally includes a block division processing unit when reading image data from the image memory unit 4. That is, the motion vector calculation unit 15 reads the image data of the target frame and the image data of the reference frame from the image memory unit 4, and the reading unit is a block unit divided into blocks.

  Then, the motion vector calculation unit 15 determines a division start point as a starting point, and reads the target block and reference block corresponding to the block Bk from the corresponding address of the image memory unit 4. That is, the motion vector calculation unit 15 reads image data from the image memory unit 4 in units of target blocks and reference blocks.

  After the motion vector detection by block matching between the target frame and the reference frame is completed, when performing block matching between the next target frame and the reference frame, the division start point of the block division is set as described in the first embodiment. So as to shift.

  Then, the motion vector calculation unit 15 performs detection processing of the local motion vector LMV, calculation of the reliability of the local motion vector LMV, calculation of the global motion, and calculation of the global motion vector GMV as described later for each block. . Further, the motion vector calculation unit 15 also evaluates the degree of background matching for the target block, and generates a hit rate β.

  In this embodiment, the motion vector calculation unit 15 outputs target frame image data TGv, reference frame image data REFv, and a hit rate β. Further, the motion vector calculation unit 15 outputs information on the NR processing motion vector MVnr composed of the global motion vector GMV or the local motion vector LMV.

  The information on the NR processing motion vector MVnr from the motion vector calculation unit 15 and the image data REFv of the reference frame are supplied to the motion compensation image generation unit 16. The motion compensated image generation unit 16 performs a process corresponding to the global motion on the image data REFv of the reference frame using the NR processing motion vector MVnr, that is, a deformation process with translation, rotation, and enlargement / reduction, A compensation image is generated.

  Then, the image data TGv of the target frame from the motion vector calculation unit 15 is supplied to the addition unit 17, and the image data MCv of the motion compensation image from the motion compensation image generation unit 16 is supplied to the addition unit 17.

  Further, the target frame image data TGv and the hit rate β from the motion vector calculation unit 15 are supplied to the addition rate calculation unit 21. The addition rate calculation unit 21 calculates the addition rate α for each target block of the target frame according to the hit rate β that is the degree of background matching. Then, the addition rate calculation unit 21 supplies the calculated addition rate α for each target block to the addition unit 17.

  The adder 17 adds the pixels at the corresponding positions of both the image data TGv and MCv to perform an image superimposition process, and outputs the image data MIXv of the added image (which is an NR image). This image superposition is performed for each target block, and the image data of the corresponding block of the image data MCv of the motion compensated image is added to the image data of the target block at the addition rate α.

  The image data MIXv of the added image from the adding unit 17 is overwritten (overwritten) on the first frame memory 41 of the image memory unit 4 as the image data of the target frame over the image data of the previous target frame.

  That is, the image data of the target frame in the first frame memory 41 is initially image data of the first frame immediately after the shutter button is pressed. When the image data MCv of the motion compensation image of the second reference frame and the target frame are added, the image data of the target frame in the first frame memory 41 is rewritten to the image data MIXv of the added image as a result of the addition. It is done.

  Then, the image data MIXv of the added image becomes the image data of the target frame with respect to the image data of the third reference frame. Then, when the image data is read from the image memory unit 4 by the motion vector calculation unit 15, the block division position is changed from the one frame before as described in the first embodiment. . That is, the division start point of the block division is changed to an offset position different from that at the time of processing in the previous frame. Then, with respect to the image data in units of blocks whose block division positions have been changed, the motion vector calculation unit 15 calculates the NR processing motion vector MVnr in the same manner as described above, and the addition unit 17 performs the image superimposition processing. The

  Then, the image data MIXv of the addition image as a result of the addition is overwritten (overwritten) in the first frame memory 41 of the image memory unit 4 as the image data of the target frame on the image data of the previous target frame. Thereafter, the same processing operation is performed for the fourth and subsequent reference frames after the division start point is further changed.

  Therefore, after the image superimposition processing for the Nth reference frame is performed, the first frame memory 41 of the image memory unit 4 has the NR in which all the N frames to be superimposed are superimposed. An image is written.

  Then, the image data MIXv of the added image, which is the NR image resulting from the superposition, stored in the first frame memory 41 of the image memory unit 4 is supplied to the still image codec unit 18 through the system bus 2 and subjected to codec conversion. . Output data from the still image codec unit 18 is recorded on a recording medium such as a DVD (Digital Versatile Disc) or a hard disk of the recording / reproducing device unit 5. In this embodiment, the still image codec unit 18 performs image compression encoding processing on a still image of JPEG (Joint Photographic Experts Group) system.

  In this still image shooting mode, before the shutter button is pressed, the image data from the data conversion unit 14 is supplied to the resolution conversion unit 19 through the first frame memory 41 of the image memory unit 4. The The image data is converted into a predetermined resolution by the resolution conversion unit 19 and then supplied to an NTSC (National Television System Committee) encoder 20. In the NTSC encoder 20, the image data is converted into an NTSC standard color video signal and supplied to a monitor display 6 made up of, for example, an LCD (Liquid Crystal Display). On the display screen of the monitor display 6, a monitor image in the still image shooting mode is displayed on the monitor.

  The image data of the still image recorded on the recording medium of the recording / reproducing apparatus unit 5 is read in response to the reproduction start operation through the user operation input unit 3, supplied to the still image codec unit 18, and reproduced and decoded. The The reproduced and decoded still image data is supplied to the NTSC encoder 20 through a buffer memory (not shown) of the image memory unit 4. The image data converted into the NTSC standard color video signal by the NTSC encoder 20 is supplied to the monitor display 6 and a reproduced image is displayed on the display screen.

  Although not shown in FIG. 18, the output video signal from the NTSC encoder 20 can be derived to the outside through a video output terminal.

  In this embodiment, the still image codec unit 18 compresses and records the image data. However, the still image codec unit 18 may be omitted and the image data may be recorded without being compressed. .

  The motion vector calculation unit 15 and the motion compensation image generation unit 16 described above can be configured by hardware. The motion vector calculation unit 15 and the motion compensation image generation unit 16 can also be configured using a DSP (Digital Signal Processor). Furthermore, the motion vector calculation unit 15 and the motion compensation image generation unit 16 can be software processed by the CPU 1.

  Similarly, the adding unit 17 can also be configured by hardware or can be configured by using a DSP. Furthermore, the addition unit 17 can also be software processed by the CPU 1. The same applies to the still image codec unit 18.

[Description of Motion Vector Calculation Unit 15]
In this embodiment, the motion vector calculation unit 15 first detects the local motion vector LMV by performing the block matching process using the SAD value described with reference to FIGS. . However, in this embodiment, the motion vector calculation unit 15 is configured by hardware as will be described later, and the local motion vector LMV can be calculated by hierarchical block matching processing.

  Further, as described above, the motion vector calculation unit 15 of this embodiment also calculates the reliability index of each of the local motion vectors LMV.

  Furthermore, the motion vector calculation unit 15 calculates global motion using only highly reliable local motion vectors. Then, a global motion vector for each block is calculated from the calculated global motion. Further, for each target block, it is determined whether it is a background portion or a moving subject portion, and a hit rate β indicating the degree of background matching is calculated.

<Layered block matching processing>
In motion vector detection processing in general conventional block matching, the reference block is moved in the search range in units of pixels (one pixel unit or a plurality of pixels), and the SAD value for the reference block at each moving position is calculated. To do. Then, the SAD value indicating the minimum value is detected from the calculated SAD values, and the motion vector is detected based on the reference block position exhibiting the minimum SAD value.

  In this embodiment, since one frame is divided into a large number of blocks, the above block matching process is performed on the entire screen while sequentially switching the target block and the reference block. Thereby, the local motion vector LMV is calculated for all the target blocks in the target frame.

  However, in such a conventional motion vector detection process, the reference block is moved in units of pixels within the search range, so that the number of matching processes for calculating the SAD value increases in proportion to the search range to be searched. This increases the matching processing time. There is also a problem that the capacity of the SAD table increases.

  In particular, this problem is significant considering that the number of pixels in a still image and the high definition (HD) of a moving image are increasing, and the size of one image is becoming very large. In addition, the number of times the image memory is accessed via the system bus 2 increases, and there is a problem that the bus bandwidth must be increased.

  In consideration of the above, in this embodiment, first, a reduced image and an intermediate image obtained by reducing a target image (target frame) and a reference image (reference frame) are prepared. Then, hierarchical block matching is performed in which the motion vector search by block matching is performed in the order of the reduced image, the intermediate image, and the original image by reflecting the result of block matching in the previous stage in the search in the next stage.

  By performing this hierarchical block matching, the calculation amount and processing time are reduced, and the local motion vector is calculated efficiently. Here, the reduced image is referred to as a reduced surface, the intermediate image is referred to as an intermediate surface, and the original image that has not been reduced is referred to as a base surface.

  FIG. 19 is a diagram for explaining this hierarchical block matching. In the example of FIG. 19, the basal plane target frame 201 and the basal plane reference frame 301 are represented by 1 / a · 1 / b (1 / a and 1 / b are reduction magnifications, and a> 1, b> 1 To a reduced plane target frame 211 and a reduced plane reference frame 311.

  In addition, the base plane target frame 201 and the base plane reference frame 301 are reduced to 1 / b to generate the intermediate plane target frame 221 and the intermediate plane reference frame 321.

  The magnification of the reduction surface and the intermediate surface with respect to the base surface is arbitrary, but a range of 1/2 to 1/8 times (1/4 to 1/64 times in terms of the number of pixels) is appropriate. In the example of FIG. 19, the reduction ratio of the reduction surface with respect to the intermediate surface is 1/4 (that is, a = 4), and the reduction magnification of the intermediate surface with respect to the base surface is 1/4 (that is, b = 4). Shown as a case.

  The method for creating the reduced surface and the intermediate surface is arbitrary. However, in the method of creating a reduced surface and an intermediate surface by simply thinning out the pixels of the original image according to the reduction magnification, a aliasing component occurs, and the motion vector detected in the first layer (reduced surface) is It becomes easy to come off from the right thing. For this reason, normally, a low-pass filter having a cutoff frequency band corresponding to the reduction magnification is applied to the original image, and then subsampling corresponding to the reduction magnification is performed.

  In this embodiment, a luminance average value including pixels that disappear due to sub-sampling according to the magnification is generated, and is used as a reduction plane pixel or an intermediate plane pixel. That is, for 1 / a reduction, the average luminance value of the square area of a × a pixels is calculated and used as the luminance value of the reduction plane pixels and intermediate plane pixels. In the case of this method, even if an intermediate surface is first generated and then a reduced surface is generated from the intermediate surface, the same result as that when the reduced surface is directly generated from the original image is obtained, which is more efficient.

  In obtaining a reduced image, the horizontal reduction ratio and the vertical reduction ratio may be the same as described above, or may be different.

  When the reduction plane and the intermediate plane are created as described above, first, the reduction plane target block 212 is set in the reduction plane target frame 211, and the reduction plane search range 313 is set in the reduction plane reference frame 311.

  Then, with respect to a plurality of reduced surface reference blocks 312 in the reduced surface search range 313, the reduced surface motion vector detection device 401 performs the above-described block matching processing to detect a reduced surface reference block position exhibiting the minimum SAD value. To do. Based on the detection of the reduced plane reference block position, the reduced plane motion vector MVs is detected.

  In this example, the motion vector detection apparatus 401 executes processing using a block having the size of the reduction plane target block 212 (the number of pixels in the horizontal direction × the number of lines in the vertical direction) as a block matching processing unit.

  When the calculation of the reduced surface motion vector MVs is completed, the intermediate surface target block 222 is set in the intermediate surface target frame 221 that is equal to the reduced surface target frame 211 multiplied by a.

  In the example of FIG. 19, the intermediate plane motion vector detection device 402 performs block matching processing using a block having the same size as the block matching processing unit in the reduced plane motion vector detection device 401 as an intermediate plane target block. . Here, the blocks of the same size have the same number of pixels and are composed of the same number of pixels in the horizontal direction × the same number of lines in the vertical direction.

  In this example, since the reduction plane is 1 / a of the intermediate plane, a region of the intermediate plane target frame corresponding to the reduction plane target block 212 includes a intermediate plane target blocks 222. Accordingly, all a of the intermediate plane target blocks 222 are set as block matching processing targets in the intermediate plane motion vector detection device 402.

  Then, in the intermediate plane reference frame 321 that is equal to the reduced plane reference frame 311 multiplied by a, an intermediate plane search range 323 centered on the reduced plane motion vector MVs is set. For the plurality of intermediate plane reference blocks 322 in the intermediate plane search range 323, the motion vector detection device 402 performs the above-described block matching process to detect the intermediate plane reference block position exhibiting the minimum SAD value. A surface motion vector MVm is detected.

  The intermediate plane motion vector detection device 402 executes block matching processing for each of the a intermediate plane target blocks in the search range for each intermediate plane target block set in the intermediate plane search range 323. Thereby, the motion vector for each intermediate plane target block is detected. A motion vector exhibiting the minimum SAD value among the plurality of motion vectors is detected as a motion vector (intermediate surface motion vector) MVm on the intermediate surface.

  When the calculation of the reduced plane motion vector MVs is completed, the base plane target block 202 is set in the base plane target frame 201 that is equal to the intermediate plane target frame 221 multiplied by b.

  In the example of FIG. 19, the motion vector detection device 403 on the base plane also processes blocks having the same size as the motion vector detection devices 401 and 402 (the same number of pixels = the same number of pixels in the horizontal direction × the same number of lines in the vertical direction). Block matching processing is performed as a unit block.

  As described above, the intermediate plane motion vector MVm is obtained in units of processing unit blocks. Therefore, the basal plane target block 202 in the basal plane target frame 201 that is the target of the motion vector detection device 403 is a block (processing unit) having the same size as the reduced plane target block, as shown by hatching in FIG. B) times the number of blocks).

  On the other hand, in the base plane reference frame 301 that is equal to the intermediate plane reference frame 321 multiplied by b, a base plane search range 303 centered on the combined vector of the reduced plane motion vector MVs and the intermediate plane motion vector MVm is set. The base vector reference block 302 in the base surface search range 303 is subjected to the above-described block matching processing by the motion vector detection device 403, and the base surface reference block position exhibiting the minimum SAD value is detected, whereby the base A surface motion vector MVb is detected.

  The reduced surface motion vector MVs and the intermediate surface motion vector MVm are obtained in units of processing unit blocks having the same size. Therefore, the set basal plane search range 303 centered on the combined vector of the reduced plane motion vector MVs and the intermediate plane motion vector MVm is a slightly wider area than the area including the b base plane target blocks 202. The

  The motion vector detection device 403 executes block matching processing in the search range for each base plane target block set in the base plane search range 303 for the b base plane target blocks 202. Thereby, the motion vector is detected for each basal plane target block. Then, a motion vector exhibiting the minimum SAD value among the plurality of motion vectors is detected as a motion vector (base motion vector) MVb on the base surface.

  Then, as a combined vector of the reduced plane motion vector MVs, the intermediate plane motion vector VMm, and the base plane motion vector VMb obtained as described above, between the base plane target frame 201 and the base plane reference frame 301, A local motion vector LMV for the basal plane target block is detected.

  The hierarchical block matching process as described above is executed in the entire region of the target frame and the reference frame while sequentially switching the target block and the reference block. As a result, all of the plurality of local motion vectors LMV of the plurality of target blocks set in the target frame are calculated.

  In the example of FIG. 19, the motion vector detection devices 401, 402 and 403 are actually one device, and are read from the target block read from the image memory unit 4 and the search range. Only the reference block is different.

  Note that the target block is switched as follows when local motion vectors LMV for all base plane target blocks 202 are obtained in the base plane target frame 201. That is, on the reduction plane, the reduction plane target block is set to be shifted in the horizontal direction by the number of pixels in the horizontal direction according to the reduction magnifications 1 / a and 1 / b. The vertical direction may be set so as to be shifted by the number of lines in the vertical direction corresponding to the reduction ratios 1 / a and 1 / b.

  However, for the purpose of obtaining the global motion from a plurality of local motion vectors LMV, the following may be performed. That is, the reduction plane target block may be set to be sequentially shifted in the horizontal direction and the vertical direction so as to obtain the local motion vector LMV for the base plane target block at the jump position in the base plane target frame 201. Good.

  Note that the above-described hierarchical block matching may be performed in two layers including only the reduced surface and the base surface, omitting the intermediate surface, and conversely, a plurality of intermediate surface layers may exist. However, care should be taken when the reduction ratio is too large and the moving subject and the background are included in the same unit block. That is, what should be detected as a different motion vector is handled as one motion vector, and recovery is not effective in the subsequent layers. Therefore, it is necessary to carefully select a reduction ratio.

[Calculation of reliability of local motion vector LMV]
When a relatively noisy image is a target image, the SAD value is affected by noise and a correct vector cannot be obtained in many cases. FIG. 20 is a photograph of a night view and is a relatively noisy image. When the motion vector between the two images is illustrated with reference to an image captured by shaking leftward with slight rotation with respect to the image of FIG. 20, the result shown in FIG. 21 is obtained. . FIG. 21 is a drawing of a reduced surface motion vector in a reduced image obtained by reducing the original image to 1/8.

  As can be seen from FIG. 21, the motion vector of the night sky, in which the texture is not particularly clear, is obtained as a disjoint motion. In the hierarchical block matching, as described above, since a low-pass filter is applied when generating a reduced image, the noise resistance is relatively high, but it is affected by noise as shown in FIG.

  Since the local motion vector LMV of the basal plane is obtained by searching around the reduced plane motion vector, if the reduced plane motion vector deviates from the correct one, the recovery is not effective and the influence of noise is decent. It will be further disturbed.

  Even if the target is a captured image with no noise, if the image is not clear, the change in the gradation due to the slight change in external light and the difference in exposure time during continuous shooting is larger. In many cases, the detected motion vector deviates from the correct one. In addition, many artificial buildings such as trees and buildings have many texture repeating patterns, but even in the case of such texture repeating patterns, the detected motion vectors are likely to deviate from the correct ones.

  Assuming such a case, as described above, conventionally, attempts have been made to calculate global motion using only highly reliable motion vectors. For example, there is a proposal that edge detection is performed on a target image and a motion vector of a block with a clear edge is highly reliable. There has also been a proposal for calculating the reliability using the DC component and AC component of the IDCT (Inverse Discreet Cosine Transform) result of the target image.

  There is also a proposal that feature points on a target image are detected in advance using a corner detector, which is a kind of filter, and their motion vectors are highly reliable. Furthermore, a method for extracting a highly reliable motion vector from a combination of motion vectors at a plurality of points has been proposed on the premise that the positional relationship between a plurality of feature points is maintained even on a reference image.

  However, none of these conventional techniques are premised on high-noise images, and it is clear that any technique fails for images with a very high noise level.

  In this embodiment, in view of the above situation, a reliability index value that can effectively evaluate the reliability of a motion vector is obtained even in an image in a high noise environment.

  In this embodiment, the difference or ratio between the first maximum value and the second maximum value of the correlation value between the target block and the reference block is used as an index value for the reliability of the motion vector. In this embodiment, since the correlation value between the target block and the reference block is detected as the SAD value, the first maximum value and the second maximum value of the correlation value are the first minimum value and the second minimum value of the SAD value. Value.

  FIG. 22 is a diagram schematically showing each SAD value in the SAD table for one target block. In FIG. 22, the search range is represented in two dimensions in the horizontal direction (x direction) and the vertical direction (y direction) of the image, and SAD values are taken in the height direction (directions orthogonal to the x direction and the y direction). The SAD table is represented by a cubic surface.

  In the case of normal block matching processing, in order to detect a motion vector, only the minimum value of the SAD value in the SAD table is detected. The minimum value of the SAD value is the first minimum value of the SAD value in the SAD table, and is the position indicated by a point 501 in FIG. In FIG. 22, the motion vector MV is detected as a vector between the motion origin (x = 0, y = 0) and the minimum position of the SAD value indicated by this point 501.

  Considering an ideal state where no noise exists, when the correlation value between a plurality of reference blocks and a target block within the search range is obtained, the SAD table represented by the cubic surface is uniformly convex downward, Only one minimum value of the SAD value exists. However, in an actual shooting situation, the SAD table represented by the cubic surface is almost uniformly convex downward due to various noises in addition to the influence of the change in the light amount and the movement of the moving object. In general, there are a plurality of minimum values of the SAD values.

  Therefore, in this embodiment, the motion vector MV is detected based on the position of the reference block that exhibits the first minimum value equal to the minimum value of the SAD value. Of the SAD values excluding the first minimum value of the SAD value, The minimum value, i.e., the second minimum value of the SAD value, is detected to generate a reliability index. In FIG. 22, the position indicated by the point 501 indicates the first minimum value, and the position indicated by the point 502 indicates the second minimum value.

  If the influence of noise or the like is small, the difference between the first minimum value of the SAD value and the second minimum value of the SAD value is large, and the movement detected from the first minimum value of the SAD value, that is, the minimum value of the SAD value. The reliability of the vector MV is increased. On the other hand, in a noisy environment, the difference between the first minimum value of the SAD value and the second minimum value of the SAD value is small, and it is not known which corresponds to the motion vector MV correctly. Reliability will be low.

  From the above, in this embodiment, the difference between the first minimum value of the SAD value (minimum value of the SAD value) and the second minimum value of the SAD value is used as an index of the reliability of the detected motion vector. FIG. 23 shows a SAD table in which the search range of FIG. 22 is represented on the one-dimensional axis. In this embodiment, in FIG. 23, the difference value between the second minimum value and the first minimum value (minimum value of the SAD value) is set as an index value Ft of the motion vector MV.

  If only the first minimum value of the SAD value is obtained and the second minimum value is not obtained, in this embodiment, the theoretical maximum value of the SAD value or the maximum value of the SAD value in the SAD table is calculated. The reliability index value of the motion vector MV is set. Therefore, although the motion vector of such a block is considered to be highly reliable, since there are almost no such blocks, only the first minimum value of the SAD value can be obtained, and the second minimum value cannot be obtained. Such block motion vectors may be excluded from the reliability evaluation.

  Instead of the difference between the first minimum value of the SAD value (minimum value of the SAD value) and the second minimum value of the SAD value, the first minimum value of the SAD value (minimum value of the SAD value) and the SAD value The ratio with the second minimum value can be used as the reliability index value Ft of the motion vector MV. However, in the following description, the difference between the first minimum value of the SAD value (minimum value of the SAD value) and the second minimum value of the SAD value is used as the motion vector reliability index value Ft. To do.

  According to the reliability index of the motion vector of this embodiment, since the image component such as the edge and the feature of the image as in the conventional case is not used, only the correlation value between the target frame and the reference frame is used. Robustness against noise. That is, a highly accurate motion vector reliability index can be obtained without being affected by image noise.

  In this embodiment, the difference or ratio between the first maximum value of the correlation value (the first minimum value of the SAD value) and the second maximum value of the correlation value (the second minimum value of the SAD value) is used. This is also why the motion vector reliability index of this embodiment is highly robust to noise.

  That is, generally, when the noise level increases, even if the motion vector is correct, the value of the SAD value increases. For this reason, when a threshold value is set for the motion vector reliability index value Ft for the purpose of extracting a highly reliable motion vector and the comparison process is performed with the threshold value, the threshold value itself is also set to the noise level. It needs to be changed accordingly.

  On the other hand, in the case of the reliability index value Ft of the motion vector according to this embodiment, when the noise level increases, the first maximum value of the correlation value (the first minimum value of the SAD value) and the second maximum value of the correlation value. Both of the values (second minimum values of the SAD values) increase according to the noise level. For this reason, the influence of noise is offset with respect to the difference between the first maximum value of the correlation value (the first minimum value of the SAD value) and the second maximum value of the correlation value.

  That is, it is possible to perform threshold processing of a constant value that does not depend on the noise level. When the ratio between the first maximum value of the correlation value (the first minimum value of the SAD value) and the second maximum value of the correlation value (the second minimum value of the SAD value) is used as the reliability index value Ft of the motion vector The same can be said for.

  By the way, when the contrast of the image of the target block to be subjected to block matching is low, the difference between the second minimum value of the SAD value and the minimum value of the SAD value tends to be reduced. For this reason, when there are a region with high contrast and a region with low contrast in the same frame, if the evaluation value Ix of the vector reliability is evaluated with the same threshold, a region with high contrast is likely to be preferentially extracted.

  This is a correct result from the viewpoint of motion vector reliability, but for the purpose of relieving a low-contrast region to some extent, in this embodiment, the arithmetic expression for calculating the motion vector reliability index value has an influence of contrast. Add a term to reduce That is, the difference between the maximum luminance value and the minimum luminance value of the image of the target frame is obtained, and the luminance difference is reflected in the motion vector reliability index value. In order to avoid adverse effects due to noise, the maximum luminance and the minimum luminance are extracted after a low-pass filter is applied to the image data of the target frame.

Based on the above, the calculation formula of the motion vector reliability index value Ft in this embodiment is:
Ft = (Btm2SAD−MinSAD) − (MaxTAR−MinTAR) × Co
... (Formula a)
However,
Ft; motion vector reliability index value Btm2SAD; second minimum value of SAD value MinSAD; minimum value of SAD value (first minimum value)
MaxTAR; maximum luminance value of target block MinTAR; minimum luminance value of target block Co; weight coefficient (≦ 1)
It becomes.

  Even when the ratio between the first maximum value of the correlation value and the second maximum value of the correlation value is used as the motion vector reliability index value, the reliability index value calculation formula is completely different from the above (formula a). Similarly, a term that reduces the influence of contrast may be added. However, when calculating the reliability index value Ft of the motion vector, it is not essential to add a term that reduces the influence of contrast, and this term may be omitted.

  In the above description, only the motion vector reliability index value for the basal plane motion vector MVb is obtained, but the motion vector reliability index value is similarly applied to the reduced plane motion vector MVs and the intermediate plane motion vector MVm. Needless to say, you can ask.

[Calculation of global motion and global motion vector GMV]
Conventionally, since there was no motion vector reliability index value as described above, global motion is calculated using all of the plurality of local motion vectors LMV obtained for the target frame with the same weight. . On the other hand, in this embodiment, the reliability index value Ft of each of the plurality of local motion vectors LMV for the target frame can be obtained as described above.

  Accordingly, the reliability index values Ft of the plurality of obtained local motion vectors LMV are normalized, and for example, a weight coefficient of 0 or more and 1 or less can be set for each local motion vector LMV. Each local motion vector LMV is not used with the same weight, but with a weight corresponding to each weight coefficient, so that the global motion can be calculated. That is, a convergence operation for calculating a global motion is started using all of the obtained plurality of local motion vectors LMV. At this time, each local motion vector LMV is determined according to the reliability index value Ft. The weight corresponding to the weight coefficient is used.

  However, in this embodiment, the weight coefficient W for the local motion vector LMV is binarized to 0 and 1 in order to simplify the calculation processing of global motion and reduce the calculation load.

For this reason, in this embodiment, a threshold th is set for the reliability index value Ft of the motion vector,
When Ft> th, W = 1
W = 0 when Ft ≦ th.
... (Formula b)
The weighting coefficient W for each local motion vector LMV is calculated by using the following formula, using the reliability index value Ft of each motion vector.

  That is, in this embodiment, the reliability of each of the plurality of local motion vectors LMV is determined using the motion vector reliability index value Ft, and the highly reliable local motion vector is determined from the plurality of local motion vectors LMV. Extract only LMV. Then, global motion is calculated using only the extracted local motion vector LMV with high reliability.

  In this embodiment, since the number of target blocks in the target frame is relatively large, accurate global motion calculation is possible even with a method that extracts only the reliable local motion vector LMV as in this example. It is.

  A specific processing example for calculating global motion from a plurality of local motion vectors LMV will be described later.

  It has been described that the local motion vector as shown in FIG. 21 is obtained for the noisy image shown in FIG. For the local motion vectors shown on the image of FIG. 21, only those whose reliability equal to or higher than the threshold is obtained after determining the reliability using the reliability index value of the motion vector according to this embodiment. Is extracted, and the block and motion vector are drawn as shown in FIG. It can be said that these blocks shown in FIG. 23 are not affected by noise, and almost correct local motion vectors are obtained.

[Hardware Configuration Example of Motion Vector Calculation Unit 15]
As described above, the motion vector calculation unit 15 detects the local motion vector LMV for each target block, calculates the reliability index value of the detected local motion vector LMV, calculates the global motion GM and the global motion vector GMV, etc. Process.

  A hardware configuration example of the motion vector calculation unit 15 is shown in FIG. In this example, the motion vector calculation unit 15 includes a target block buffer unit 151 that holds pixel data of the target block 102 and a reference block buffer unit 152 that holds pixel data of the reference block 108.

  In addition, the motion vector calculation unit 15 includes a matching processing unit 153 that calculates SAD values for pixels corresponding to the target block 102 and the reference block 108. In addition, the motion vector calculation unit 15 includes a local motion vector calculation unit 154 that calculates a local motion vector from the SAD value information output from the matching processing unit 153. The motion vector calculation unit 15 further includes a control unit 155, a motion vector reliability index value calculation unit 156, a global motion calculation unit 157, a global motion vector calculation unit 158, a contrast calculation unit 159, and a background / moving subject. A determination unit 150.

  The contrast calculation unit 159 includes a low-pass filter 1591, a maximum luminance value detection unit 1592, and a minimum luminance value detection unit 1593.

  In this example, although not shown, the image memory unit 4 stores the reduced plane image data and the intermediate plane image data for the target frame and the reference frame, which are generated from the target frame and reference frame image data of the original picture. Is retained.

  The control unit 155 controls the processing sequence in the motion vector calculation unit 15, and supplies a control signal to each unit as shown in the figure.

  Under the control of the control unit 155, the target block buffer unit 151 reads the image data of the instructed target block from the image data of the target frame on the reduced surface, intermediate surface, or base surface of the image memory unit 4, This is supplied to the matching processing unit 153.

  Under the control of the control unit 155, the reference block buffer unit 152 reads image data in the designated matching processing range from the image data of the reference frame on the reduced surface, intermediate surface, or base surface of the image memory unit 4. Then, the reference block buffer unit 152 sequentially supplies the image data of the reference block to the matching processing unit 153 from the image data in the matching processing range.

  When the control unit 155 finishes the processing in units of blocks for one set of target frames and reference frames, the target block buffer unit 151 and the reference block buffer unit 152 are configured to shift the division start point of block division as described above. To control. Based on the control of the control unit 155, the target block buffer unit 151 and the reference block buffer unit 152 read the target block and the reference block from the image memory unit 4 based on the division start point shifted from the previous frame.

  The matching processing unit 153 receives the target block image data from the target block buffer unit 151 and the reference block image data from the reference block buffer unit 152. Then, the matching processing unit 153 performs block matching processing on the reduced surface, the intermediate surface, and the base surface according to the control of the control unit 155. Then, the matching processing unit 153 supplies the reference vector (reference block position information) and the SAD value of the block matching processing result to the local motion vector calculation unit 154.

  The local motion vector calculation unit 154 includes a first minimum value holding unit 1541 of the SAD value and a second minimum value holding unit 1542 of the SAD value, and the first minimum of the SAD value is obtained from the SAD value from the matching processing unit 153. A process of detecting the value and the second minimum value of the SAD value is performed.

  The local motion vector calculation unit 154 includes the first minimum value of the SAD value and the position information (reference vector) of the first minimum value holding unit 1541 of the SAD value, and the SAD value of the second minimum value holding unit 1542 of the SAD value. The second minimum value and its position information (reference vector) are updated. The local motion vector calculation unit 154 performs this update processing until the block matching processing for all reference blocks in the matching processing range is completed.

  When the block matching process is completed, the first minimum value holding unit 1541 of the SAD value stores and holds the first minimum value of the SAD value and the position information (reference vector) for the target block at that time. Further, the second minimum value holding unit 1542 of the SAD value stores and holds the second minimum value of the SAD value and its position information (reference vector).

  The local motion vector calculation unit 154, when the block matching process for all the reference blocks in the matching processing range is completed, information on the reference vector (position information) held in the first minimum value holding unit 1541 of the SAD value Are detected as motion vectors on each of the reduced surface, the intermediate surface, and the base surface. The processing operation of the local motion vector calculation unit 154 will be described in detail later.

  In this embodiment, the local motion vector calculation unit 154 supplies the reduced surface motion vector MVs to the control unit 155 as the local motion vector LMV during the reduced surface matching process.

  When performing the hierarchical block matching process, the control unit 155 determines the search range on the intermediate plane from the information of the reduced plane motion vector MVs. Then, the control unit supplies control signals to the target block buffer unit 151, the reference block buffer unit 152, and the matching processing unit 153, and controls to perform block matching on the intermediate plane.

  When the matching process of the intermediate plane is completed, the local motion vector calculation unit 154 supplies information on the combined vector of the reduced plane motion vector MVs and the intermediate plane motion vector MVm to the control unit 155 as the local motion vector LMV. .

  The control unit 155 determines a search range in the base plane from information on a combined vector of the reduced plane motion vector MVs and the intermediate plane motion vector MVm. Then, the control unit 155 supplies a control signal to the target block buffer unit 151, the reference block buffer unit 152, and the matching processing unit 153 so as to perform block matching on the base surface.

  As described above, the controller unit 155 can control up to which layer of the hierarchical block matching is performed and the motion vector in the layer is output as the local motion vector LMV.

  For example, when it is desired to obtain the local motion vector LMV in a short time without sacrificing accuracy, the local motion vector calculation unit 154 may perform only the reduced surface block matching processing. When a highly accurate local motion vector LMV is required, the local motion vector calculation unit 154 may perform a hierarchical block matching process up to a basal plane block matching process. Furthermore, in consideration of accuracy and time, the local motion vector calculation unit 154 may be better to perform a hierarchical block matching process up to an intermediate plane block matching process.

  When the layered matching processing up to the base plane is completed, the local motion vector calculation unit 154 receives information on the combined vector of the reduced plane motion vector MVs, the intermediate plane motion vector MVm, and the base plane motion vector MVb. Obtained as LMV. The local motion vector LMV is supplied to the global motion vector calculation unit 157.

  The global motion calculation unit 157 temporarily stores the received local motion vector LMV.

  When the local motion vector calculation unit 154 completes the local motion vector LMV calculation process, the motion vector reliability index value calculation unit 156 is enabled by the control unit 155. Further, from the local motion vector calculation unit 154, the minimum Sin value MinSAD of the first minimum value holding unit 1541 and the second minimum value Btm2SAD of the SAD value of the second minimum value holding unit 1542 are motion vector reliability. The index value calculation unit 156 is supplied.

  At this time, the image data of the target block from the target block buffer unit 151 is supplied to the maximum luminance value detection unit 1592 and the minimum luminance value detection unit 1593 through the low pass filter 1591. Then, the maximum brightness value MaxTAR and the minimum brightness value MinTAR detected by the maximum brightness value detection unit 1592 and the minimum brightness value detection unit 1593 are supplied to the motion vector reliability index value calculation unit 156.

  The motion vector reliability index value calculation unit 156 calculates the reliability index value Ft of the motion vector according to the above-described (formula a) using the information supplied thereto. Then, the motion vector reliability index value calculation unit 156 supplies the calculated motion vector reliability index value Ft to the global motion calculation unit 157. The global motion calculation unit 157 temporarily holds the reliability index value Ft of the input motion vector in association with the local motion vector LMV supplied at that time.

  When the above processing is completed for all target blocks of the target frame, the control unit 155 supplies the global motion calculation unit 157 with a control instruction signal for starting the global motion calculation processing.

  In this embodiment, the global motion calculation unit 157 receives the control instruction signal, and first, for the plurality of local motion vectors LMV held, the corresponding motion vector reliability index value Ft is stored. To determine reliability. In this embodiment, the weight coefficient W for each local motion vector LMV is calculated by the above-described (formula b). The global motion calculation unit 157 extracts only the reliable local motion vector LMV with the weight coefficient W = 1.

  Then, a calculation process for calculating the global motion GM is executed using only the extracted local motion vector LMV with high reliability. Then, the global motion calculation unit 157 supplies the calculated global motion GM to the global motion vector calculation unit 158.

  The global motion vector calculation unit 158 calculates the global motion vector GMV of the target block by applying the global motion GM to the coordinate position (for example, the center position) of the target block.

  The global motion vector calculation unit 158 supplies the calculated global motion vector GMV to the background / moving subject determination unit 150. The background / moving subject determination unit 150 is also supplied with the local motion vector LMV from the local motion vector calculation unit 154.

  As described above, the background / moving subject determination unit 150 compares the local motion vector LMV for each target block with the global motion vector GMV, and determines the degree of coincidence between the target blocks, that is, the degree of background coincidence. judge. In this case, the background / moving subject determination unit 150 compares the correlation values (SAD values) for the reference blocks corresponding to the local motion vector LMV and the global motion vector GMV, and determines the background and the moving subject. Do.

  By the way, the local motion vector LMV and the correlation value (SAD value) obtained in order to calculate the global motion by the local motion vector calculation unit 154 can be used for the comparison in the background / moving subject determination unit 150.

  However, in that case, the local motion vector calculation unit 154 needs to hold the local motion vector and the SAD value for the processing time in the global motion calculation unit 157 and the global motion vector calculation unit 158. In this case, in particular, since the SAD value to be held does not know to which reference vector the global motion vector GMV corresponds, the entire SAD value table is held for each target block. It is necessary to keep. For this reason, a memory having a huge capacity is required as a memory for holding these local motion vectors LMV and SAD values.

  In view of this, in this embodiment, the local motion vector calculation unit 154 recalculates the comparison local motion vector LMV and SAD value in the background / moving subject determination unit 150. Thereby, the local motion vector calculation unit 154 does not need to be provided with a memory for holding the local motion vector LMV and the SAD value, and the problem of the capacity of the memory can be avoided.

  Therefore, the background / moving subject determination unit 150 uses the recalculated local motion vector LMV and SAD value to obtain a hit rate β indicating the degree of background matching for the target block. The SAD value for the reference vector (reference block position) that matches the global motion vector GMV is also acquired at the time of the recalculation. Then, the background / moving subject determination unit 150 determines whether the target block is a background portion or a moving subject portion using the recalculated local motion vector LMV and SAD value.

  As described above, the background / moving subject determination unit 150 compares the SAD value for the reference block corresponding to the global motion vector GMV with the SAD value for the reference block corresponding to the local motion vector LMV. Correct the value to reflect the minutes.

  Image noise generally depends on the luminance value of the image. Therefore, in this embodiment, the maximum luminance value MaxTAR from the maximum value luminance value detection unit 1592 of the contrast calculation unit 159 and the minimum luminance value MinTAR from the minimum luminance value system unit 1593 are supplied to the background / moving subject determination unit 150. Is done.

  The background / moving subject determination unit 150 detects the luminance value of the image as a difference between the maximum luminance value MaxTAR and the minimum luminance value MinTAR, and the reference block corresponding to the global motion vector GMV according to the detected luminance value. The correction value of the SAD value is determined.

  Then, the background / moving subject determination unit 150 compares the corrected correlation value with the correlation value for the reference block corresponding to the local motion vector. Then, it is determined whether the correction value of the SAD value for the reference block corresponding to the global motion vector GMV is smaller than the correlation value for the reference block corresponding to the local motion vector (the correlation value is large). When the target block is small, the target block is evaluated as a background portion.

  The background / moving subject determination unit 150 outputs the global motion vector GMV as the NR processing motion vector MVnr when the hit rate β is such that the target block can be regarded as the background portion. In other cases, the background / moving subject determination unit 150 outputs the local motion vector LMV as the NR processing motion vector MVnr to the target block.

  As described above, the NR processing motion vector MVnr from the background / moving subject determination unit 150 is supplied to the motion compensation image generation unit 16.

  In the motion compensation image generation unit 16, the deformation processing corresponding to the motion vector MVnr for NR processing is performed on the image data REFv of the reference frame sent through the motion vector calculation unit 15, and the motion compensation image data MCv. Is generated. Then, the generated motion compensation image data MCv is supplied to the adding unit 17.

  The background / moving subject determination unit 150 supplies the obtained hit rate β to the addition rate calculation unit 21. The addition rate calculation unit 21 calculates the addition rate α of the motion compensated image with respect to the target image based on the hit rate β.

  In this case, the addition rate α can be calculated in units of blocks or pixels. Even when the calculation is performed in units of pixels, the hit rate β is in units of blocks, so it goes without saying that the hit rate β is involved in the calculation of the addition rate in units of blocks.

  When the addition rate α is calculated in units of blocks, the addition rate α corresponding to only the hit rate β can be calculated. Of course, the addition rate α can also be calculated by adding other addition rate calculation elements other than the hit rate β. In this embodiment, the addition rate α is calculated not only by the hit rate β but also by taking into account factors such as a difference in pixel units between the target image and the motion compensated image and noise of the target image, as will be described later.

  The addition rate α obtained by the addition rate calculation unit 21 is supplied to the addition unit 17. The adding unit 17 adds the image data MCv of the motion compensated image with the addition rate α to the image data TGv of the target image sent thereto.

  In this embodiment, the adder 17 superimposes the target image and the motion compensated image at an addition rate according to the background matching degree (hit rate β) of the target block. That is, the addition rate of the motion compensation image to the target image is high in the still image portion, and the addition rate of the motion compensation image to the target image is low in the moving subject portion.

  Thus, in this embodiment, the background portion and the moving subject portion can be overlapped with each other at an appropriate addition rate. Therefore, the NR image obtained by image superposition is good.

  In addition, since the global motion and the global motion vector GMV in this embodiment are generated from the highly reliable local motion vector LMV, they are highly accurate motion vectors, and a good NR image can also be obtained in that respect. Can do.

[Flow of overall processing for NR image generation]
FIG. 26 is a flowchart showing a flow of processing for generating an NR image for a captured image taken into the image memory 4. Note that FIG. 26 shows processing for a set of target frames and reference frames.

  First, the control unit 155 of the motion vector calculation unit 15 sets the division start point for the frame image to be processed as described in the first embodiment, and the block division position is the block in the previous frame. It decides so that it may not overlap with a division position (Step S1).

  Next, in the motion vector calculation unit 15, the local motion vector calculation unit 154 detects the local motion vector LMV by block matching (step S2). Next, the motion vector reliability index value calculation unit 156 uses the reliability index value Ft for the detected local motion vector LMV as the first minimum value MinSAD of the SAD value calculated by the local motion vector calculation unit 154. The second minimum value Btm2SAD is used for calculation (step S3).

  Next, the global motion calculation unit 157 extracts a highly reliable local motion vector LMV based on the calculated reliability index value Ft. Then, the global motion calculation unit 157 calculates the global motion GM using only the extracted local motion vector LMV with high reliability (Step S4).

  Next, a global motion vector GMV for each target block is calculated using the global motion GM calculated in step S3 (step S5).

  Further, the local motion vector calculation unit 154 performs re-block matching processing to re-detect the local motion vector LMV (step S6). The SAD value (MinSAD) at the reference block position (reference vector) where the local motion vector LMV is detected is held, and the SAD value for the reference block corresponding to the global motion vector GMV calculated in step S4 is also held. Keep it. At this time, there is no need to calculate the reliability index value Ft of the local motion vector LMV.

  Next, the background / moving subject determination unit 150 compares the local motion vector LMV calculated in step S6 with the global motion vector GMV calculated in step S5, thereby determining the background matching degree for each target block. Evaluate and obtain the hit rate β. Further, an NR processing motion vector MVnr for each target block is determined and output to the motion compensated image generating unit 16 (step S7).

  In the comparison between the local motion vector LMV and the global motion vector GMV in step S7, the SAD value for the reference block corresponding to the local motion vector LMV, the SAD value for the reference block corresponding to the global motion vector GMV, Compare Further, as described above, the SAD value for the reference block corresponding to the global motion vector GMV is corrected according to image noise.

  Next, the motion compensation image generation unit 16 generates a motion compensation image for each reference block using the NR processing motion vector MVnr determined in step S6, and outputs the motion compensation image to the addition unit 17 (step S8).

  Next, as described above, the addition rate calculation unit 21 calculates the addition rate α for each pixel and supplies it to the addition unit 17 as described above using the hit rate β obtained in step S7 ( Step S9).

  Next, the adding unit 17 generates an NR image by superimposing the motion compensation image generated in step S7 on the target image for each pixel at the addition rate α calculated in step S9 (step S10). This completes the processing of the NR image.

  FIG. 26 shows a processing operation for one target frame. When superposing three or more images, the process of FIG. 26 is repeated with the generated NR image as a target frame and the newly superposed image as a reference frame.

  Next, detailed processing operations of each main part of the motion vector calculation unit 15 will be further described.

[Processing Operation in Local Motion Vector Calculation Unit 154]
In the local motion vector calculation unit 154 of this embodiment, in order to detect the minimum value for the SAD value, the difference value for the target pixel (own pixel) is compared with the difference values of the pixels around the target pixel. To. That is, in FIG. 27, as indicated by the dotted line, the position Po is a determination target point (pixel), the SAD value of the determination target point position Po, and the eight pixels near the determination target point position Po. Are compared with the SAD value. Then, it is determined whether or not the SAD value of the pixel of the determination target point is the minimum value (referred to as a local minimum) among the nine SAD values in the area surrounded by the dotted line.

  When it is determined that the SAD value of the determination target point is a local minimum, whether or not it is smaller than the minimum value of the held SAD value compared to the minimum value of the held SAD value. Judge. When it is determined that the SAD value of the determination target point is smaller than the minimum value of the held SAD value, the minimum value of the held SAD value is updated to the newly detected local minimum SAD value. Go.

  As a configuration for detecting the SAD value of the local minimum as described above, in this embodiment, the local motion vector calculation unit 154 reduces the scale as a buffer memory of the SAD table that holds the SAD value. It is devised to. That is, when searching the reference block in units of one pixel, as a buffer memory of the SAD table that holds the SAD value, as shown in FIG. Prepare one that can hold the minute SAD value.

  As can be seen from FIG. 27, when the SAD values for 2 lines + 3 in the horizontal size of the target block are written in the buffer memory, the local minimum at the determination target point position Po can be determined.

  In order to minimize the size of the buffer memory, here, as shown in FIG. 27, a new input is made to the memory location Pa where the SAD value that is no longer used in the local minimum evaluation (local minimum detection) is stored. The SAD value is overwritten. That is, the newly input SAD value is written in the memory location Pb in FIG. 27 as an order, but by reusing the memory location Pa that is no longer used instead of this memory location Pb, The increase in the size of the wear is suppressed.

  In this embodiment, the local motion vector calculation unit 154 includes the first minimum value holding unit 1541 and the second minimum value holding unit 1542 described above in addition to the buffer for detecting the local minimum. is there.

  The above process is a basic process, and the minimum value of the SAD value and the second minimum value of the SAD value are detected by applying this basic process to the first minimum value and the second minimum value. Like that.

  In this embodiment, the local motion vector calculation unit 154 performs the same operation on the reduced surface, the intermediate surface, and the base surface. However, in this embodiment, the local motion vector LMV is detected on the base surface. And the reliability index value is calculated. Accordingly, in this example, the second minimum value of the SAD value is required only in the base surface, and the calculation and holding of the second minimum value of the SAD value is omitted on the reduction surface and the intermediate surface. Also good.

  FIG. 28 is a flowchart showing the flow of processing for detecting the first minimum value and the second minimum value in the local motion vector calculation unit 154.

  First, the SAD value from the matching processing unit 153 is fetched (step S101). Then, the SAD value at the determination target point position Po is compared with the surrounding eight SAD values (step S102). As a result of the comparison, it is determined whether or not the SAD value of the determination target point position Po is a local minimum (step S103).

  If it is determined in step S103 that the SAD value at the determination target point position Po is not a local minimum, the process returns to step S101, and the next SAD value is captured.

  If it is determined in step S103 that the SAD value at the determination target point position Po is a local minimum, the first minimum value and the second minimum value of the held SAD value, the SAD value at the determination target point position Po, and Are compared (step S104).

  Then, it is determined whether or not the SAD value at the determination target point position Po is smaller than the first minimum value of the held SAD value (step S105). When it is determined that the SAD value is small, the SAD value held in the first minimum value holding unit 1541 is updated to be held in the second minimum value holding unit 1542, and the SAD value at the determination target point position Po is changed to the first minimum value. The value is held in the value holding unit 1541 (step S106).

  Then, it is determined whether or not the calculation processing for the SAD values for all the reference blocks with respect to the target block has been completed (step S109). When it is determined that the calculation has not been completed, the process returns to step S101. The next SAD value is taken in. If it is determined in step S109 that the calculation processing for the SAD values between all the reference blocks is completed for the target block, this processing routine is terminated.

  When it is determined in step S105 that the SAD value at the determination target point position Po is larger than the first minimum value of the held SAD value, the SAD value at the determination target point position Po is held at the held SAD value. It is determined whether it is smaller than the second minimum value (step S107). When it is determined that the SAD value is small, the SAD value held in the second minimum value holding unit 1542 is updated to the SAD value of the determination target point position Po (step S108).

  Then, after step S108, the process proceeds to step S109, where it is determined whether or not the calculation process for the SAD values between all the reference blocks is completed for the target block. When it is determined that the process has not been completed, the process returns to step S101 to capture the next SAD value. If it is determined in step S109 that the calculation processing for the SAD values between all the reference blocks is completed for the target block, this processing routine is terminated.

  The process also proceeds to step S109 when it is determined in step S107 that the SAD value at the determination target point position Po is not smaller than the second minimum value of the held SAD value. In step S109, it is determined whether or not the calculation process for the SAD values between all the reference blocks for the target block has been completed.

  In step S109, when it is determined that the calculation processing for the SAD values between all the reference blocks is not completed for the target block, the process returns to step S101 to capture the next SAD value. If it is determined in step S109 that the calculation processing for the SAD values between all the reference blocks is completed for the target block, this processing routine is terminated.

  The processing flowchart of FIG. 28 can be the same for the reduction surface, the intermediate surface, and the base surface. In this case, on the reduced surface and the intermediate surface, the SAD value finally held in the first minimum value holding unit 1541 is detected as the minimum value MinSAD of the SAD value, and the corresponding reference vector is used as the reduced surface motion vector. MVs and intermediate plane motion vectors MVm can be detected and output.

  Further, on the base surface, the SAD value finally held in the first minimum value holding unit 1541 is detected as the minimum value MinSAD of the SAD value, and the corresponding reference vector is detected as the base surface motion vector MVb. Can be output. In addition, on the basal plane, the SAD value (minimum value MinSAD) finally held in the first minimum value holding unit 1541 and the SAD value (the first SAD value of the SAD value) finally held in the second minimum value holding unit 1542 2 minimum value) is supplied to the motion vector reliability index value calculation unit 156.

[Processing in Global Motion Calculation Unit 157]
<About expanded affine transformation>
In this embodiment, global motion is calculated (estimated) from a large number of local motion vectors LMV. Then, a global motion vector GMV is calculated (estimated global motion) from the calculated global motion. In this case, a method of expressing global motion by affine transformation is used.

  FIG. 29 shows a general formula of affine transformation as (Formula 1). In this (Formula 1), v represents the horizontal component of the motion vector of the target block, w represents the vertical component of the motion vector of the target block, and a, b, c, d, e, and f represent affine parameters. Yes. Here, in normal affine transformation, the affine parameters a, b, c, d, e, and f are fixed values. X and y indicate the horizontal and vertical components of the center coordinates of the target block.

  A motion vector corresponding to the global motion (this vector is called an ideal motion vector) is obtained from the affine parameters obtained in the global motion convergence calculation process and the coordinates of the center of each target block. The total error ε between this ideal vector and the observed motion vector (detected by block matching) is expressed as (Equation 2) in FIG.

  The proposition of deriving the global motion is the estimation of the affine parameters a to f that minimize the total error ε described above, and this can be solved by, for example, the least square method. 31, 32, and 33, (Equation 3), (Equation 4), and (Equation 5) show the derivation process of the affine parameters a to f and the results thereof.

  Thus, although the affine transformation parameters are relatively easily calculated, the effect is great. Affine transformation can cope with a certain amount of deformation in addition to parallel movement, rotation, and enlargement / reduction of an image, and thus covers most camera shake, that is, slight correction of camera work.

  As a prior art document on the above affine transformation, there is Patent Document 2 (Japanese Patent Laid-Open No. 2005-321902).

  However, the normal affine transformation has a drawback that it cannot cope with “aori”. The “tilting” is a phenomenon caused by the rotational component of the pitch axis (vertical axis) or the yaw axis (horizontal axis perpendicular to the vertical direction) of camera shake. In addition, it is also called trapezoidal distortion or keystone because it deforms into a trapezoid.

  That is, as shown on the lower side of FIG. 34A, when the camera CAM faces the subject OBJ so that the optical axis Lz of the camera CAM is orthogonal to the rectangular plane of the subject OBJ, the rectangle of the subject OBJ The captured image of the plane is a rectangular plane as it is as shown in the upper side of FIG.

  On the other hand, for example, as shown on the lower side of FIG. 34B, the optical axis Lz of the camera CAM does not face the object OBJ, but rotates by the pitch axis by the angle θ (rotates in a vertical plane). In this case, the captured image of the rectangular plane of the subject OBJ is a trapezoid whose length in the left-right direction (the horizontal direction of the image) changes linearly according to the angle θ, as shown in the upper side of FIG. It becomes.

  Although illustration is omitted, when the optical axis Lz of the camera CAM does not face the subject OBJ and rotates the yaw axis by the angle θ (rotates in the horizontal plane), the captured image of the rectangular plane of the subject OBJ Is a trapezoid whose length in the vertical direction (vertical direction of the image) changes linearly according to the angle θ.

  The process of returning the image captured as described above to the original shape by camera shake is “tilting” correction or trapezoidal correction, but this cannot be expressed by affine transformation parameters. That is, for example, when considering the “tilt” correction for returning the “tilt” trapezoidal image as shown in FIG. 34B to the rectangle, the top is horizontally reduced along the vertical axis (y-axis), and the bottom is Just zoom in horizontally. However, of the parameters for affine transformation, the parameter for horizontal enlargement / reduction is a, and since this parameter a is a constant value, affine transformation cannot cope with this correction.

  In order to avoid such a problem, it is considered to use orthogonal projection transformation instead of affine transformation. The orthographic transformation is expressed by (Equation 6) in FIG. 35, and can represent a projection transformation from an arbitrary plane to a plane in a three-dimensional space.

  However, it is extremely difficult to apply the orthographic transformation in a case where one global motion is specified for a large number of vectors as assumed here. This is because it is difficult to use the method of least squares from the shape of (Equation 6) in FIG.

  As a practical problem, since the actual imaging target is a three-dimensional object, the observed local motion vector includes a large number of subjects having different distances. In that case, the orthographic transformation fit is just an approximation. Therefore, considering the enormous cost required to solve the general solution of orthographic transformation, it can be said that it is reasonable to judge that the cost effectiveness is not balanced.

  In the case where orthographic projection transformation is applied more simply, a method for obtaining a solution for six vectors can be employed. Since there are 12 parameters to be obtained, if 6 vector coordinates are substituted, 6 × 2 sets of 6-ary linear equations can be obtained, which can be solved relatively easily by inverse matrix calculation of 6 × 6 matrix. It is because it can be derived. Therefore, it is only necessary to appropriately select six vectors from a large number of local vectors.

  However, considering the case where the captured image includes a moving subject, each of the local vector accuracy cannot be expected so much, so it is difficult to extract only a highly accurate vector from among many local motion vectors. It is. For this reason, it is very difficult to appropriately select six vectors from a large number of local motion vectors.

  In this embodiment, in view of the above points, a technique is used that allows a global motion to be detected satisfactorily even if the image has an “aori” deformation.

  That is, in this embodiment, an extended affine transformation (hereinafter referred to as an extended affine transformation) obtained by modifying the conventional normal affine transformation shown in FIG. 29 is used.

  Considering the “tilt” correction in which the “tilt” trapezoidal image as shown by the solid line in FIG. 36 is returned to the rectangle (rectangle) shown by the dotted line in FIG. 36, the vertical axis (y Along the axis, the upper half of the trapezoidal image may be horizontally reduced and the lower half may be horizontally enlarged.

  In the normal affine transformation, among the parameters a, b, c, d, e, and f of the affine transformation, the parameter relating to horizontal enlargement / reduction is a, and since the parameter a is a constant value, in the normal affine transformation, Cannot handle “Aori” correction.

  Therefore, as shown in the lower side of FIG. 36, if the parameter a is replaced with q0 · y + r0 so that the value changes along the y-axis, the horizontal enlargement / reduction ratio changes linearly according to the vertical axis. Will be converted. That is, the parameter a is replaced with a function of the y-axis variable y for “tilting” correction. Here, in order to correct the “tilt” in FIG. 36, the linear function, that is, the parameter “a” may be a linear function q0 · y + r0.

  FIG. 36 considers the correction of the “tilt” accompanying the rotation of the pitch axis described above, but the correction of the “tilt” accompanying the rotation of the yaw axis and the rotation of both the pitch axis and the yaw axis are combined. It is necessary to consider the correction of “Aori”.

  Therefore, in this embodiment, the idea described with reference to FIG. 36 is extended. That is, in order to expand and reduce the affine transformation along an arbitrary axis in a direction perpendicular to the axis, parameters a, b, d, and e are set as shown in (Expression 7) in FIG. Pn · x + qn · y + rn (n = 0, 1, 2, 3).

  Here, the affine parameters a, b, d, and e are parameters related to the deformation of the image, whereas the affine parameters c and f are parameters related to the horizontal and vertical shifts of the image. Therefore, the affine parameters c and f are not replaced with functions.

  Expressing this as (Equation 7), there are a total of 14 parameters, and the extended affine transformation seems to be complicated, but if (Equation 7) is expanded and rearranged, 12 parameters are eventually obtained. By using this, the shape of (Equation 8) in FIG. 38 is obtained.

  Although this extended affine transformation formula shape is complex, it is possible to uniquely derive solutions for a plurality of local motion vectors by the method of least squares. Since the method is the same as the affine transformation, only the final result is shown as (Equation 9 to (Equation 13)) in FIGS.

  In this embodiment, as shown in FIG. 37, all of the affine parameters a, b, d, e involved in the deformation of the image are replaced with functions. However, the deformation of the image is performed only in a specific direction. In some cases, only the parameters related to the direction need to be replaced with a function. For example, when only the deformation shown in the upper part of FIG. 36 is considered, only the parameter a may be replaced with a function as shown in the lower part of FIG.

  In the above example, the replacement function is assumed to be a linear function because a linear displacement is assumed. However, when a curved displacement is assumed, it may be a quadratic or higher function.

<Calculation of global motion using expanded affine transformation>
In this embodiment, the method shown in the flowcharts of FIGS. 44 and 45 is used as a method for deriving the most suitable global motion from an unspecified number of vectors. This method uses the least-squares method to converge global motion parameters while gradually eliminating local motion vectors of blocks that are difficult to adapt to global motion, such as moving subjects, from among highly reliable blocks. It is a method to plan.

  In the processing of the flowcharts of FIGS. 44 and 45, by introducing extended affine transformation, error vectors such as moving subjects are excluded from many local motion vectors with low vector accuracy, and “origin” is also included. Thus, it is possible to derive the optimum global motion at a realistic calculation cost.

  By the way, since the extended affine transformation of this embodiment has many deformation parameters and is flexible, it corresponds to an erroneous motion vector such as a moving subject or noise in the global motion convergence calculation. The vector may not be excluded.

  Therefore, in this embodiment, as shown in FIGS. 44 and 45, in the initial stage of the convergence calculation loop, an erroneous motion vector (referred to as an error vector) is eliminated using a normal affine transformation, and then an extended affine is performed. A convergence operation using transformation is performed. This is a method of trying to achieve high-precision convergence using the extended affine transformation after the vector error has been reduced to the extent that the normal affine transformation is not used to eliminate the vector of the tilt component. This is because of using.

  In this embodiment, a vector error as a difference between a motion vector (global motion vector) obtained from the global motion GM obtained at each round of the convergence calculation and a detected motion vector (the above-described local motion vector LMV). The maximum value of is detected.

  Then, when the detected maximum value of the vector error is larger than a predetermined threshold value, normal affine transformation is continued, and when the detected maximum value of the vector error becomes equal to or smaller than the predetermined threshold value, the extended affine transformation is performed. A convergence operation using transformation is performed.

  Description will be made along the flowcharts of FIGS. 44 and 45.

  First, the motion vector calculation unit 157 compares each motion vector reliability index value Ft with a predetermined threshold value for a plurality of local motion vectors LMV held. Then, from the comparison result, only the target block of the local motion vector LMV with a high reliability whose motion vector reliability index value Ft is larger than a predetermined threshold (hereinafter referred to as a block for the sake of simplicity) is selected. (Step S201). This process corresponds to the case where binary values 1 and 0 are used as the weighting coefficient W as described above using (Equation b).

  Next, the motion vector calculation unit 157 determines whether or not the convergence loop of the convergence calculation is the first time (step S202). When it is determined that this is the first time, the motion vector calculation unit 157 uses only the local motion vectors LMV of the plurality of selected blocks and derives (estimates) the global motion GM using normal affine transformation. (Step S203). That is, the affine parameters a to f for the global motion GM are calculated.

  Next, based on the derived global motion GM, the theoretical local motion vector LMVs of each of the selected blocks used in the calculation is calculated (step S206).

  Next, in each of the selected plurality of blocks, an error En between the local motion vector LMV obtained by the block matching process and the theoretical local motion vector LMVs obtained in step S206 is calculated (step S207).

  The error calculation between the motion vector obtained by the block matching and the theoretical motion vector may be performed by correctly calculating the distance from the three-square theorem if the calculation accuracy is important. However, if lightness is more important than calculation accuracy, both horizontal and vertical distances may be obtained and their sum may be used as an approximate distance.

  Next, using all of the errors En obtained for each of the selected blocks, an average value Eave and a maximum value Emax of all the errors are calculated (step S208). Then, it is determined whether or not the average value Eave is smaller than a predetermined threshold value θa and the maximum value Emax is smaller than a predetermined threshold value θb (step S209).

  As a result of the determination in step S209, when it is determined that the condition is not satisfied, a plurality of blocks for deriving the global motion are determined for the block having the error En = Emax from the error En of each block obtained in step S207. Exclude from the block. Alternatively, a block having an error En ≧ θb is detected, and all of the detected blocks are excluded from a plurality of blocks from which the global motion GM is derived (step S211 in FIG. 45).

  Next, as a result of block exclusion in step S211, it is determined whether or not the number of remaining blocks is smaller than a predetermined number of threshold values θc (step S212). When it is determined in step S212 that the number of remaining blocks is not less than the threshold value θc, the process returns to step S202, and the processes after step S202 are repeated with the remaining blocks selected.

  When the number of remaining blocks is smaller than the threshold value θc, an appropriate global motion GM cannot be obtained, so that the target reference frame image cannot be used for image superimposition in this embodiment. Therefore, when it is determined in step S212 that the number of remaining blocks is smaller than the threshold value θc, the process skips not to perform all subsequent processing for the reference frame (step S213).

  Next, when it is determined in step S202 that the convergence loop of the convergence calculation is not the first time, the motion vector calculation unit 157 determines the maximum value Emax of each block error En determined in step S207 in advance. It is determined whether or not it is larger than the threshold value θd (step S204).

  Here, the threshold value θd is a value that does not exclude the tilt component vector when the global motion GM is calculated using normal affine transformation and error vectors are eliminated as in step S211 described above. Selected.

  When it is determined in step S204 that the maximum value Emax of the error En is larger than the predetermined threshold value θd, the motion vector calculation unit 157 proceeds to step S203 and derives the global motion GM using normal affine transformation. Like that. That is, a parameter for extended affine transformation is calculated. Then, the processes after step S203 described above are repeated.

  In step S204, when it is determined that the maximum value Emax of the error En is equal to or less than the predetermined threshold θd, the motion vector calculation unit 157 uses only the local motion vectors LMV of the selected plurality of blocks and expands them. Global motion is derived using affine transformation (step S205). Then, after step S205, the processing after step S206 described above is repeated.

  Then, when it is determined in step S206 in FIG. 44 that the average value Eave of the error En is smaller than the threshold value θa and the maximum value Emax of the error En is smaller than the threshold value θb, it is assumed that the calculation has converged and the global motion GM Are determined at that time (step S210). Then, this processing routine ends.

  In step S211, whether only blocks with an error En of the maximum error Emax are excluded or blocks with an error En equal to or greater than the threshold value θb are collectively determined according to the speed of convergence and accuracy in obtaining the global motion vector GMV. You can decide by balance. If priority is given to accuracy, the former method may be adopted, and error blocks may be eliminated one by one. If priority is given to the speed of convergence, the latter method may be selected.

  In the above-described example, in step S204, the maximum value Emax of the error En is used for determination of switching between the calculation using the normal affine transformation and the calculation using the extended affine transformation. However, the switching determination in step S204 described above may consider not only the maximum value Emax of the error En but also the average value Eave of the error En.

  That is, in step S204, it is determined whether or not the average value Eave of the error En is larger than a predetermined threshold value. When the average value Eave is larger than the threshold value, a calculation using normal affine transformation is performed. Then, when the average value Eave becomes equal to or less than the threshold value, the calculation using the extended affine transformation is started.

  In step S204, it is determined whether or not both the maximum value Emax of the error En and the average value Eave are larger than a predetermined threshold value. When it is determined that the value is larger than the threshold value, an operation using normal affine transformation may be performed, and when it is determined that the value is smaller than the threshold value, calculation using extended affine transformation may be started.

  In addition, since the range of camera shake correction can be identified from the optical zoom magnification and the set shape, the maximum value of the tilt component can be obtained from the maximum amount of camera shake, and therefore the above-described threshold value identification is relatively easy.

  The global motion vector calculation unit 157 calculates a global motion vector GMV for each target block based on the global motion GM calculated as described above. That is, using the calculated affine transformation parameters a to l for the global motion GM (see FIG. 38), the motion vector (theoretical local motion vector LMVs) for each target block is obtained by (Equation 8) in FIG. Equivalent). In (Equation 8) in FIG. 38, the center position coordinates of each target block are used for x and y. The obtained motion vector becomes the global motion vector GMV for each target block.

  Then, the global motion vector GMV for each target block obtained as described above is supplied to the motion compensation image generation unit 16. Then, the motion compensation image generation unit 16 generates a motion compensation image for each target block using the global motion vector GMV for each target block, and supplies the generated motion compensation image for each block to the addition unit 17. .

  In this embodiment, the global motion vector for each target block obtained for the noisy image of FIG. 20 is shown in FIG. In FIG. 46, the global motion vector for each target block is indicated by a white line arrow.

[Configuration Example and Processing Operation of Background / Moving Subject Determination Unit 150]
In the background / moving subject determination unit 150, as described above, the SAD value for the reference block corresponding to the global motion vector GMV calculated by the global motion vector calculation unit 158 is re-block matched by the local motion vector calculation unit 154. Get at the time of.

  In this re-block matching, the block division position changing process is not performed.

  Then, the acquired SAD value is corrected with the correction value according to the noise of the image as described above. In this embodiment, two methods are provided as correction methods using the correction values.

<First Example: Correction of SAD Value by Adding Offset>
FIG. 47 is a diagram for explaining the idea of the first example. FIG. 47 is a diagram showing the contents (SAD values) of the SAD table for one target block when the horizontal axis represents the search range and the vertical axis represents the SAD value. Each value on the horizontal axis is a reference block position (reference vector), and a solid line indicates the contents of the SAD table. This is exactly the same as that shown in FIG.

  In FIG. 47, the position (that is, the reference vector) 501 of the reference block having the minimum SAD value is detected as a local motion vector LMV by block matching, as in FIG. On the other hand, it is assumed that the position of the reference block serving as the global motion vector GMV is a position 503 in FIG.

  At this time, if the SAD value in the local motion vector LMV and the SAD value in the global motion vector GMV are within a difference range corresponding to the noise amount of the image, the global motion vector GMV is a reference vector that takes the minimum SAD value. It may have become.

  In other words, the SAD value in the global motion vector GMV (reference block position) should be minimized, but the position of a different reference block (this is the local motion vector LMV) is detected as the minimum due to noise. It can be said that it has been done.

Therefore, in this example, the offset value OFS corresponding to the noise of the image is added to the SAD value in the global motion vector GMV for correction. In this example, the offset value OFS is corrected by subtracting it from the SAD value (referred to as SAD_GMV) in the global motion vector GMV. That is, if the corrected SAD value is MinSAD_G,
MinSAD_G = SAD_GMV-OFS (Formula c)
It becomes.

  Here, if the offset value OFS is set too large, even the motion vector of the moving subject is determined as the minimum SAD value. Therefore, in this embodiment, the offset value OFS is set to a value smaller than the threshold th (see formula b) for the reliability index value Ft of the local motion vector.

  The background / moving subject determination unit 150 compares the corrected SAD value MinSAD_G with the SAD value (MinSAD) in the local motion vector LMV. As a result of the comparison, if MinSAD_G <MinSAD, the minimum value of the SAD value for the target block is evaluated as MinSAD_G, which is a correction value of the SAD value for the reference block corresponding to the global motion vector GMV. FIG. 47 shows a case where MinSAD_G <MinSAD.

  That is, as shown in FIG. 47, when MinSAD_G <MinSAD, it is determined that the true local motion vector for the target block matches the global motion vector GMV. In that case, the background matching degree for the target block is evaluated to be high, and the hit rate β for the target block is set to a large value. The NR processing motion vector for the target block is a global motion vector GMV.

  Here, in this embodiment, the hit rate β is set to a value in the range of 0 ≦ β ≦ 1. When the local motion vector LMV completely matches the global motion vector GMV, the hit rate β can be set to β = 1. However, in this example, it is not evaluated whether the local motion vector LMV and the global motion vector GMV coincide with each other, and the corrected SAD value MinSAD_G and the SAD value (MinSAD) in the local motion vector LMV are Evaluate by comparison. Therefore, the evaluation result includes a case where the local motion vector LMV and the global motion vector GMV completely match, but also includes a case where they do not match.

  As will be described later, in this embodiment, if MinSAD_G <MinSAD, the hit rate β is set to β = 0.75, for example. Here, β is not set to 1 because, as described above, the comparison result includes a case where the local motion vector LMV and the global motion vector GMV do not match.

  FIG. 48 shows a configuration example (functional block diagram) of the background / moving subject determination unit 150 in the case of the first example, along with its peripheral parts.

  In the example of FIG. 48, the background / moving subject determination unit 150 includes a SAD_GMV detection unit 1501, an offset subtraction unit 1502, an offset generation unit 1503, and a comparison determination unit 1504.

  In this example, the SAD value calculated at the time of re-block matching from the matching processing unit 153 and the position information (reference vector information) of each reference block are supplied to the local motion vector calculation unit 154. The local motion vector calculation unit 154 detects the minimum SAD value MinSAD at the time of re-block matching, and calculates a local motion vector LMV (reference block position information (reference vector information)). Then, the local motion vector calculation unit 154 supplies the local motion vector LMV obtained as a result of the re-block matching process and the minimum SAD value MinSAD to the comparison determination unit 1504.

  The SAD value and the position information (reference vector information) of each reference block calculated at the time of re-block matching from the matching processing unit 153 are also supplied to the SAD_GMV detection unit 1501. The SAD_GMV detection unit 1501 detects the SAD value SAD_GMV for the reference block of the reference vector that matches the global motion vector GMV, and supplies the detected SAD value SAD_GMV to the offset subtraction unit 1502.

  The offset subtraction unit 1502 subtracts the offset value OFS from the offset generation unit 1503 from the SAD value SAD_GMV from the SAD_GMV detection unit 1501 to generate a corrected SAD value MinSAD_G. Then, the offset subtraction unit 1502 supplies the generated corrected SAD value MinSAD_G to the comparison determination unit 1504.

  The offset generation unit 1503 generates an offset value OFS according to the image noise and supplies the offset value OFS to the offset subtraction unit 1502. In view of the fact that the image noise depends on the luminance value of the image, in this example, information of the correspondence table between various luminance values and the offset value OFS is stored in the offset generation unit 1503.

  The offset generation unit 1503 calculates the luminance value of the target block as a difference between the maximum luminance value MaxTAR from the maximum luminance value detection unit 1592 and the minimum luminance value MinTAR from the minimum luminance value detection unit 1593. Then, the offset generation unit 1503 searches the correspondence table using the calculated luminance value of the target block as an argument, and acquires an offset value OFS corresponding to the luminance value of the argument.

  As described above, the comparison determination unit 1504 compares the minimum SAD value MinSAD corresponding to the local motion vector LMV with the corrected SAD value MinSAD_G from the offset subtraction unit 1502, and according to the comparison result, the hit rate β Is generated and output.

  The comparison determination unit 1504 is also supplied with the recalculated local motion vector LMV from the local motion vector calculation unit 154 and the global motion vector GMV from the global motion vector calculation unit 158.

  When MinSAD_G <MinSAD, the comparison determination unit 1504 outputs the global motion vector GMV as the NR processing vector MVnr to the motion compensation image generation unit 16 because the degree of background matching is high. When MinSAD_G <MinSAD is not satisfied, the comparison determination unit 1504 outputs the local motion vector LMV as the NR processing vector MVnr to the motion compensation image generation unit 16 because the degree of background matching is low.

  Note that the SAD value detection unit 1501 and the offset subtraction unit 1502 of the global motion vector GMV may be configured to be included in the local motion vector calculation unit 154. In that case, the local motion vector calculation unit 154 detects the SAD value SAD_GMV for the reference block corresponding to the global motion vector GMV while performing re-block matching, and performs the subtraction operation of the offset OFS.

<Flow of background / moving subject determination processing>
The re-block matching process in the local motion vector calculation unit 154 and the background matching degree evaluation process in the background / moving subject determination unit 150 configured as shown in FIG. 48 will be further described with reference to the flowcharts in FIGS. 49 and 50. . 49 and 50 are executed by each unit based on the control of the control unit 155. 49 and FIG. 50 is a case where the SAD_GMV detection unit 1501 and the offset subtraction unit 1502 are configured to be included in the local motion vector calculation unit 154.

  First, re-block matching is started under the control of the control unit 155, and the first target block is set in the target block buffer unit 151 (step S301). Next, under the control of the control unit 155, a reference block for block matching processing is set from the image data of the reference frame in the matching processing range buffered in the reference block buffer 152 (step S302).

  The matching processing unit 153 performs block matching processing on the set target block and the set reference block, and calculates an SAD value. Then, the matching processing unit 153 supplies the obtained SAD value to the local motion vector calculation unit 154 together with the position information (reference vector) of the reference block (step S303).

  The local motion vector calculation unit 154 determines whether or not the reference vector matches the global motion vector GMV (step S304). This corresponds to the processing operation of the SAD_GMV detection unit 1501.

  If it is determined in step S304 that the reference vector does not match the global motion vector GMV, the local motion vector calculation unit 154 performs an update process on the minimum SAD value MinSAD and its reference block position (reference vector). (Step S305). That is, the minimum SAD value MinSAD held so far is compared with the newly calculated SAD value, the smaller SAD value is held as the minimum SAD value MinSAD, and the reference block position (reference vector) is also set. Update to the one that exhibits the minimum SAD value.

  Next, the control unit 155 determines whether or not block matching with the target block for all reference blocks within the search range has been completed (step S308).

  If it is determined in step S308 that the processing for all reference blocks in the search range has not been completed, the control unit 155 sets the next reference block (step S309). And it returns to step S302 and repeats the process after this step S302 mentioned above.

  When it is determined in step S304 that the reference vector matches the global motion vector GMV, the local motion vector calculation unit 154 subtracts the offset value OFS from the SAD value SAD_GMV (step S306). Then, the subtraction result is used as the corrected SAD value MinSAD_G and the reference block position (reference vector = global motion vector GMV) is held (step S307).

  Then, the process proceeds to step S308, where it is determined whether or not the processing for all reference blocks within the search range has been completed.

  If it is determined in step S308 that the processing for all reference blocks within the search range has been completed, the local motion vector calculation unit 154 detects the local motion vector LMV and the minimum SAD value MinSAD, and compares and determines 1504 is supplied. The local motion vector calculation unit 154 also supplies the corrected SAD value MinSAD_G to the comparison determination unit 1504 (step S311 in FIG. 50).

  Then, the comparison / determination unit 1504 of the background / moving subject determination unit 150 determines whether or not the minimum SAD value MinSAD is smaller than a predetermined threshold value TH1 (step S312). The processing in step S312 is provided to cope with the following problem.

  That is, for example, when a block corresponding to a moving subject portion on the target image disappears due to movement on the reference image or is accompanied by a large deformation, any reference block in the search range can be matched with the target block. No situation arises.

  In such a case, the SAD value in the SAD table is a large value and the same value, and the minimum SAD value MinSAD and other values are close to each other. In such a situation, the corrected SAD value MinSAD_G obtained by subtracting the offset value OFS from the SAD value at the reference vector that matches the global motion vector GMV is always smaller than the minimum SAD value MinSAD, and is erroneously detected as a background portion. There is a risk that

  Therefore, in order to avoid this problem, in the present embodiment, the process of determining the background matching degree by correcting the SAD value in the reference vector that matches the global motion vector GMV is the minimum SAD value MinSAD is the threshold TH1. If it is larger, do not do it.

  Accordingly, when it is determined in step S312 that the minimum SAD value MinSAD is larger than the threshold value TH1, the comparison determination unit 1504 does not determine the background matching degree using the global motion vector GMV, and the hit rate β = 0. 25 (step S313). The reason why β = 0 is not considered is that there may be a block that matches the background among target blocks that exhibit an SAD value greater than the threshold TH1.

  Then, the comparison determination unit 1504 outputs the recalculated local motion vector LMV to the motion compensation image generation unit 16 as the NR processing motion vector of the target block (step S314).

  Next, the control unit 155 determines whether or not processing has been completed for all target blocks in the target frame (step S320), and if not, sets the next target block (step S321). And it returns to step S302 and repeats the process after step S302 mentioned above under control of the control part 155. FIG.

  If it is determined in step S312 that the minimum SAD value MinSAD is smaller than the threshold value TH1, the comparison determination unit 1504 compares the minimum SAD value MinSAD with the corrected SAD value MinSAD_G (step S315). Then, the comparison determination unit 1504 determines whether or not MinSAD> MinSAD_G (step S316). If it is determined that this is not the case, the target block does not match the background, and the hit rate β is β = It is set to 0 (step S317).

  After step S317, the process proceeds to step S314, and the comparison determination unit 1504 outputs the recalculated local motion vector LMV to the motion compensation image generation unit 16 as the NR processing motion vector of the target block. And it progresses to step S320 and repeats the process after step S320 mentioned above.

  If it is determined in step S316 that MinSAD> MinSAD_G, the hit rate β is set to β = 0.75, assuming that the target block is highly consistent with the background (step S318). Then, the comparison determination unit 1504 outputs the global motion vector GMV to the motion compensation image generation unit 16 as the NR processing motion vector of the target block (step S319).

  After step S319, the process proceeds to step S320, and the control unit 155 determines whether or not the processing for all target blocks in the target frame has been completed. If it is determined in step S320 that the processing has not been completed for all target blocks in the target frame, the control unit 155 sets the next target block (step S321), and then returns to step S302.

  If it is determined in step S320 that the processing for all target blocks in the target frame has been completed, the control unit 155 ends the background / moving subject determination processing operation.

  In the above-described embodiment, the offset value OFS is subtracted from the SAD value SAD_GMV of the global motion vector GMV to obtain the corrected SAD value MinSAD_G, which is compared with the minimum SAD value MinSAD. However, even if the value (MinSAD + OFS) obtained by adding the offset value OFS to the minimum SAD value MinSAD of the local motion vector LMV is compared with the SAD value SAD_GMV of the global motion vector GMV, they are completely equivalent.

<Second Example: Correction of SAD Value by Gain Multiplication>
In the first example described above, the SAD value SAD_GMV for the reference block at the reference vector position matching the global motion vector GMV is corrected by adding the offset value OFS. Based on the same concept, the SAD value SAD_GMV can be corrected by multiplying the gain g considering the noise of the image instead of adding the offset. In this example, since the SAD value SAD_GMV is multiplied by a gain in a smaller direction, the gain g is g <1. Of course, g> 1 when the correlation value increases in accordance with the strength of the correlation.

In this second example, the corrected SAD value MinSAD_G of the SAD value SAD_GMV in the global motion vector GMV is:
MinSAD_G = SAD_GMV × g (formula d)
It becomes.

  FIG. 51 shows a configuration example (functional block diagram) of the background / moving subject determination unit 150 in the case of the second example, along with its peripheral parts.

  In the example of FIG. 51, in the configuration of the example of FIG. 48 described above, a gain multiplication unit 1505 is provided instead of the offset subtraction unit 1502, and a gain generation unit 1506 is provided instead of the offset generation unit 1503. Become.

  The gain generation unit 1506 stores the information of the correspondence table between various luminance values and gain g (g <1) since the noise of the image corresponds to the luminance value of the image, and the luminance value of the target block And the corresponding gain g is output.

  The gain multiplication unit 1505 multiplies the gain g from the gain generation unit 1506 by the SAD value SAD_GMV from the SAD_GMV detection unit 1501 to generate a corrected SAD value MinSAD_G. Then, the gain multiplication unit 1505 supplies the generated corrected SAD value MinSAD_G to the comparison determination unit 1504. Other configurations and processing operations are the same as those in the first example described above.

  In the background / moving subject determination process in the second example, the correction process of the SAD value SAD_GMV in step S306 in the flowchart of FIG. 49 is changed to the above-described calculation process of (formula d). The processing of the other steps can be made exactly the same as in the first example.

  In the above-described embodiment, the SAD value SAD_GMV of the global motion vector GMV is multiplied by the gain g (g <1) to obtain the corrected SAD value MinSAD_G and compared with the minimum SAD value MinSAD. did. However, even if a value obtained by multiplying the minimum SAD value MinSAD of the local motion vector LMV by the reciprocal of the gain g (MinSAD × 1 / g) and the SAD value SAD_GMV of the global motion vector GMV are completely equivalent.

  Note that the gain g in the second example is also a value that avoids being determined as the minimum SAD value up to the motion vector of the moving subject, like the offset OFS in the first example. That is, in this embodiment, similarly to the offset value OFS, the gain g is also set to a value that takes into consideration the threshold value th (see formula b) for the local motion vector reliability index value Ft.

[Other Configuration Examples and Processing Operation of Background / Moving Subject Determination Unit 150]
By the way, the global motion vector GMV calculated based on the global motion GM can be originally expressed with a finer accuracy than the pixel pitch, whereas in the above-described embodiment, the block matching is performed with the accuracy of the pixel pitch. Done. For this reason, an error may occur in the reference block position of the reference vector that matches the global motion vector GMV.

  Moreover, even if the affine transformation expanded as in the above-described embodiment is used, there may be an error between the obtained global motion GM and the true global motion. In addition, in the hierarchical block matching in this embodiment, if a mistake occurs in shallow layer block matching, the mistake cannot be recovered in the subsequent layers. Considering the above, it can be said that it is better to provide a certain margin with respect to the position of the global motion vector GMV.

  The example described below is an example in consideration of the above. In this example, not only the SAD value at the position Pgmv of the global motion vector GMV but also the SAD value at the vicinity of the position is corrected as described above and compared with MinSAD.

  For example, as shown in FIG. 52, it is assumed that the position of the global motion vector GMV is determined to be the position Pgmv in the SAD table TBL for the target block. At this time, not only the SAD value of the position Pgmv but also the eight SAD values of the neighboring neighboring positions P1 to P8 adjacent to the position Pgmv by one pixel are corrected as described above to match the background. It is a target for determining the degree.

  In this example, the corrected SAD values MinSAD_G_0 to MinSAD_G_8 of the nine SAD values at the position Pgmv and the positions P1 to P8 are compared with the recalculated minimum SAD value MinSAD. If any of the nine corrected SAD values is smaller than the minimum SAD value MinSAD, the motion vector of the target block is evaluated as having a high degree of coincidence with the global motion vector GMV, and the degree of background is somewhat high. Evaluate as high.

  In this example, FIG. 53 shows a configuration diagram around the background / moving subject determination unit 150 when the first example described above is used as a method of correcting the SAD value.

  As shown in FIG. 53, in this example, instead of the SAD_GMV detection unit 1501 in FIG. 48, a GMV neighborhood SAD value detection unit 1507 is provided, and a minimum value is provided between the offset subtraction unit 1502 and the comparison determination unit 1504. In this configuration, a detection unit 1508 is provided.

  The GMV vicinity SAD value detection unit 1507 includes the position Pgmv of the global motion vector GMV and its SAD value, and the eight positions P1 to P8 around the position, from the position information and SAD value from the matching processing unit 153. The SAD value is detected.

  Then, the GMV vicinity SAD value detection unit 1507 supplies the detected SAD value to the offset subtraction unit 1502. The offset subtraction unit 1502 subtracts the offset value OFS from the offset generation unit 1503 for each of the nine SAD values at the positions Pgmv and P1 to P8, and generates corrected SAD values MinSAD_G_0 to MinSAD_G_8. Then, the offset subtraction unit 1502 supplies the corrected SAD values MinSAD_G_0 to MinSAD_G_8 to the minimum value detection unit 1508.

  Further, the GMV vicinity SAD value detection unit 1507 supplies information on the detected positions Pgmv and P1 to P8 to the minimum value detection unit 1508 so as to correspond to the SAD values at the respective positions.

  The minimum value detection unit 1508 detects the minimum value among the nine corrected SAD values MinSAD_G_0 to MinSAD_G_8 of the positions Pgmv and P1 to P8, and compares and determines the minimum value of the detection result as the MinSAD_G with the position information. Part 1504.

  In the comparison determination unit 1504, MinSAD obtained as a result of re-block matching from the local motion vector 154 is compared with MinSAD_G from the minimum value detection unit 1508 in the same manner as described above, and background matching is performed as described above. Determining the degree of evaluation.

  When MinSAD> MinSAD_G is not satisfied, the comparison determination unit 1504 sets the hit rate β to β = 0.0 and outputs the local motion vector LMV as the motion vector for NR processing. This is similar to the above example.

  When MinSAD> MinSAD_G, the comparison / determination unit 1504 determines whether or not the position information corresponding to the MinSAD_G matches the global motion vector GMV, and changes the hit rate β according to the result. For example, when the position information corresponding to MinSAD_G matches the global motion vector GMV, the hit rate β is set to β = 0.75. When the position information corresponding to MinSAD_G does not match the global motion vector GMV, the hit rate β is set to β = 0.5.

  Even in this example, when MinSAD> MinSAD_G, the global motion vector GMV is output as the motion vector for NR processing.

  However, in this example, when MinSAD> MinSAD_G, the comparison determination unit 1504 outputs the local motion vector LMV corresponding to the position information from the minimum value detection unit 1508 as a motion vector for NR processing. Good.

  The minimum value detection unit 1508 may be configured to be included in the comparison determination unit 1504. In that case, the minimum value is detected from the nine corrected SAD values and then compared with the minimum SAD value MinSAD, or the nine corrected SAD values and the minimum SAD value MinSAD are sequentially compared. Of course, it is good.

<Flow of background / moving subject determination processing>
The flow of the re-block matching process in the local motion vector calculation unit 154 and the background matching degree evaluation process in the background / moving subject determination unit 150 having the configuration of FIG. 53 will be further described with reference to the flowcharts of FIGS. . The processes in FIGS. 54 and 55 are executed by each unit based on the control of the control unit 155. 54 and 55, the SAD_GMV detection unit 1501 and the offset subtraction unit 1502 are configured to be included in the local motion vector calculation unit 154. In addition, the minimum value detection unit 1508 is included in the comparison determination unit 1504.

  First, re-block matching is started under the control of the control unit 155, and the first target block is set in the target block buffer unit 151 (step S401). Next, under the control of the control unit 155, a reference block for block matching processing is set from the image data of the reference frame in the matching processing range buffered in the reference block buffer 152 (step S402).

  The matching processing unit 153 performs block matching processing on the set target block and the set reference block, and calculates an SAD value. The matching processing unit 153 supplies the calculated SAD value to the local motion vector calculation unit 154 together with the position information (reference vector) of the reference block (step S403).

  The local motion vector calculation unit 154 determines whether the reference vector matches the global motion vector GMV and whether it matches the eight positions P1 to P8 in the vicinity thereof (step S404).

  If it is determined in step S404 that the reference vector does not match the global motion vector GMV and the eight neighboring positions P1 to P8, the local motion vector calculation unit 154 determines the minimum SAD value MinSAD and its reference block position. Update processing with (reference vector) is performed (step S405).

  Next, the control unit 155 determines whether or not block matching with the target block for all the reference blocks within the search range has been completed (step S408).

  If it is determined in step S408 that the processing has not been completed for all reference blocks in the search range, the control unit 155 sets the next reference block (step S409). And it returns to step S402 and repeats the process after this step S402 mentioned above.

  If it is determined in step S404 that the reference vector matches the global motion vector GMV or any one of the neighboring positions P1 to P8, the local motion vector calculation unit 154 subtracts the offset value OFS from the SAD value. (Step S406). Then, the subtraction result is held as one of the corrected SAD values MinSAD_G_0 to MinSAD_G_8 and the reference block position (reference vector) is held (step S407).

  Then, the process proceeds to step S408, and it is determined whether or not the processing for all reference blocks within the search range has been completed.

  When it is determined in step S408 that the processing for all reference blocks within the search range has been completed, the local motion vector calculation unit 154 detects the local motion vector LMV and the minimum SAD value MinSAD, and compares and determines 1504 is supplied. Further, the local motion vector calculation unit 154 detects the corrected SAD value MinSAD_G as the minimum value among the nine corrected SAD values MinSAD_G_0 to MinSAD_G_8 and supplies the detected value to the comparison determination unit 1504 (step S411 in FIG. 55). .

  Then, the comparison / determination unit 1504 of the background / moving subject determination unit 150 determines whether or not the minimum SAD value MinSAD is smaller than the predetermined threshold TH1 (step S412).

  If it is determined in step S412 that the minimum SAD value MinSAD is larger than the threshold TH1, the comparison determination unit 1504 does not determine the background matching degree using the global motion vector GMV, and the hit rate β = 0.25. (Step S413). Then, the comparison determination unit 1504 outputs the recalculated local motion vector LMV to the motion compensation image generation unit 16 as the NR processing motion vector of the target block (step S414).

  Then, after step S414, the control unit 155 determines whether or not processing for all target blocks in the target frame has been completed (step S422), and if not, sets the next target block. (Step S423). Then, the process returns to step S402, and the processes after step S402 described above are repeated under the control of the control unit 155.

  If it is determined in step S412 that the minimum SAD value MinSAD is smaller than the threshold TH1, the comparison determination unit 1504 compares the minimum SAD value MinSAD with the corrected SAD value MinSAD_G (step S415). Then, the comparison determination unit 1504 determines whether or not MinSAD> MinSAD_G (step S416). If it is determined that the target block does not match the background, the hit rate β is determined as β = It is set to 0 (step S417).

  After step S417, the process proceeds to step S414, and the comparison determination unit 1504 outputs the recalculated local motion vector LMV to the motion compensation image generation unit 16 as the NR processing motion vector of the target block. And it progresses to step S422 and the process after step S422 mentioned above is repeated.

  If it is determined in step S416 that MinSAD> MinSAD_G, it is determined whether the position (reference vector) of MinSAD matches the global motion vector GMV (step S418). If it is determined in step S418 that the MinSAD position (reference vector) does not match the global motion vector GMV, the comparison determination unit 1504 sets the hit rate β to β = 0.5 (step S418). S419).

  If it is determined in step S418 that the MinSAD position (reference vector) matches the global motion vector GMV, the comparison determination unit 1504 sets the hit rate β to β = 0.75 (step S418). S420).

  Then, after step S419 or 420, the comparison determination unit 1504 outputs the global motion vector GMV to the motion compensated image generation unit 16 as the NR processing motion vector of the target block (step S421).

  After step S421, the process proceeds to step S422, and the control unit 155 determines whether or not the processing for all target blocks in the target frame has been completed. If it is determined in step S422 that the processing has not been completed for all target blocks in the target frame, the control unit 155 sets the next target block (step S423), and then returns to step S402.

  If it is determined in step S422 that the processing for all target blocks in the target frame has been completed, the control unit 155 ends the background / moving subject determination processing operation.

  The method of determining the background / moving subject in consideration of not only the global motion vector GMV position but also the SAD value in the vicinity of the surrounding area can be applied to the second example of the SAD value correction method described above. .

  FIG. 56 shows a configuration example of the background / moving subject determination unit and its peripheral part when the second example is applied. In the example of FIG. 56, instead of the SAD_GMV detection unit 1501 in FIG. 52, a GMV neighborhood SAD value detection unit 1507 is provided, and a minimum value detection unit 1508 is provided between the gain multiplication unit 1505 and the comparison determination unit 1504. This is a configuration provided.

  The processing operation of the example of FIG. 56 is exactly the same as the example of FIG. 38 described above, except that the SAD value correction method is changed to the second example, and detailed description thereof is omitted here.

<Configuration Example of Addition Rate Calculation Unit 21>
In this embodiment, the addition rate calculator 21 calculates the addition rate α (0 ≦ α ≦ 1) for each pixel. As described above, basically, the addition rate α is calculated according to the difference in pixel units between the target image and the motion compensated image and the hit rate β. For this reason, the hit rate β, the target image data TGv, and the motion compensation image data MCv are supplied to the addition rate calculation unit 21.

  Here, if the difference in pixel unit between the target image and the motion compensated image is large, it is considered as a moving part. Therefore, the addition rate α should be low, but if there is no difference in the pixel unit or is small, the addition rate α may be high. However, when it is unclear whether the target block is the background still image portion or the moving subject portion, as described above, the addition rate α is set so as to make the double exposure in the moving portion inconspicuous. Must be kept relatively low.

  However, in this embodiment, since the degree of coincidence as a background still image portion is obtained from the hit rate β, the addition rate α is made higher as the hit rate β is larger and the background coincidence degree is higher.

  Further, in this embodiment, since the noise of the image further depends on the brightness of the image, the addition rate calculation unit 21 determines the difference in pixel units between the target image and the motion compensated image. The noise according to the brightness is taken into consideration. That is, if the pixel unit difference between the target image and the motion compensated image is within the noise range, it is determined that the actual difference between the two is small. When the difference in pixel unit is larger than the noise range, it is determined that the difference between the two is actually large.

  In order to consider this noise, the maximum luminance value MaxTAR and the minimum luminance value MinTAR are supplied from the motion vector calculation unit 15 to the addition rate calculation unit 21. Then, the addition rate calculation unit 21 calculates the brightness for each target block, that is, the illuminance Lx, from the maximum luminance value MaxTAR and the minimum luminance value MinTAR.

  The addition rate calculation unit 21 of this embodiment is supplied with an addition adjustment gain GA so that the addition degree can be controlled from outside the addition rate calculation unit 21. The value of the addition adjustment gain GA can be configured to be selected by a user from gain values prepared in advance. Or you may comprise so that the value of the gain GA for addition adjustment may be defined according to the ISO sensitivity of a captured image. In the latter case, for example, the brightness of the image can be determined according to the exposure value at the time of imaging, the shutter speed, and the like, and the value of the addition adjustment gain GA can be determined. That is, since noise is conspicuous in the case of a bright image, the gain is set to be small so that the addition rate is small. Conversely, in the case of a dark image, the gain is set to be large so as to increase the addition rate.

  A hardware configuration example of the addition rate calculation unit 21 of this embodiment is shown in FIGS. In the example of FIGS. 57 and 58, the target image data TGv and the motion compensation image data MCv are processed separately for each of the luminance component and the color difference component.

  FIG. 57 is a processing portion for the luminance component in the addition rate calculation unit 21. With reference to FIG. 57, first, the addition rate calculation processing for the luminance component will be described.

  That is, the luminance data Ytg in the target image data TGv is supplied to the difference absolute value calculation unit 603 through the low-pass filter 601. Also, the luminance data Ymc of the motion compensation image data MCv is supplied to the difference absolute value calculation unit 603 through the low-pass filter 602. The difference absolute value calculation unit 603 calculates the pixel difference absolute value ΔPxY between the luminance data Ytg and the luminance data Ymc for each pixel, and supplies the calculated pixel difference absolute value ΔPxY to the addition rate conversion output unit 608.

  The luminance data Ytg that has passed through the low-pass filter 601 is also supplied to the spatial average filter 604. The luminance data Ymc that has passed through the low-pass filter 602 is also supplied to the spatial average filter 605.

  Each of the spatial average filters 604 and 605 is an average value Mtg and Mmc of nine pixels of the eight pixels around the pixel (referred to as a target pixel) for which a pixel difference is calculated by the difference absolute value calculation unit 603 and the own pixel. Ask for. Then, the spatial average filters 604 and 605 supply the obtained average values Mtg and Mmc to the difference absolute value calculation unit 606.

  The difference absolute value calculation unit 606 calculates an average difference absolute value ΔMeY between the average value Mtg and the average value Mmc, and supplies the calculated average difference absolute value ΔMeY to the addition rate conversion output unit 609.

  In this example, a noise model memory 607 for luminance components is provided. In the noise model memory 607, the noise standard deviation σ with respect to the luminance value of the pixel value is calculated and stored in advance. That is, when the luminance value of the pixel value is taken on the horizontal axis and the noise standard deviation σ is taken on the vertical axis, the noise standard deviation σ is as shown in FIG. 59, for example. As shown in FIG. 59, the standard deviation σ of noise is small for bright pixels, and the standard deviation σ of noise is large for dark pixels.

  Since the noise model differs depending on the brightness of the image, the noise model memory 607 stores a plurality of noise models corresponding to the illuminance Lx of the image.

  Then, as described above, the illuminance Lx for each target block obtained from the maximum luminance value MaxTAR and the minimum luminance value MinTAR is supplied to the noise model memory 607 for selection of a noise model corresponding to the illuminance. Thereby, which noise model is used is determined by the illuminance Lx.

  Also, the luminance data Ytg of the target image data TGv that has passed through the low-pass filter 601 is supplied to the noise model memory 607, and a noise standard deviation σY determined by the illuminance Lx and the luminance data Ytg is obtained from the noise model memory 607. This noise standard deviation σY is supplied to addition rate conversion output units 608 and 609.

  Further, the addition rate conversion output units 608 and 609 are supplied with a hit rate β from the motion vector calculation unit 15 and an external gain GA.

  The addition rate conversion output unit 608 uses the conversion function with the pixel difference absolute value ΔPxY, the noise standard deviation σY, the hit rate β, and the gain GA as conversion parameters, and the addition rate αYA based on the pixel difference absolute value ΔPxY. Is output.

  Also, the addition rate conversion output unit 609 performs addition based on the average difference absolute value ΔMeY using a conversion function having the average difference absolute value ΔMeY, the noise standard deviation σY, the hit rate β, and the gain GA as conversion parameters. The rate αYB is output.

  The addition rate conversion output units 608 and 609 basically determine the addition rate α from the noise standard deviation σ and the difference value as shown in FIG. In addition rate conversion output units 608 and 609, noise standard deviation σ is noise standard deviation σY from noise model memory 607, and difference values are pixel difference absolute value ΔPxY and average difference absolute value ΔMeY. The output addition rate α is the addition rates αYA and αYB.

  In this embodiment, as shown in FIG. 60, when the difference value is equal to or smaller than a predetermined first threshold value that can be set as a multiple of the noise standard deviation σ, the addition rate α is the maximum value as a background image portion. It is assumed that α = 1.

  Then, as shown in FIG. 60, the addition rate is determined as a portion where the difference value cannot be determined as either the background portion or the moving subject portion within the range of (first threshold value <difference value <second threshold value). α is linearly decreased as the difference value increases. In the range of (difference value ≧ second threshold), it is determined that the subject is a moving subject, and the addition rate α is set to α = 0, which is the minimum value.

  When the background matching degree of the target block is unknown, as described above, in order to reduce a state such as double exposure in the moving subject portion, the first threshold is set to a value that allows for safety and is added. The rate α is suppressed as a whole. That is, if it is known that it is a background image portion, the first threshold value and the second threshold value can be set as large values, and image addition can be performed at as high an addition rate as possible.

  However, when the degree of background coincidence is unknown, it is necessary to consider reducing the state of double exposure in the moving subject portion for all target blocks. Therefore, normally, as shown in the example of FIG. 60, the first threshold value is set to about 1 time (1σ) of the standard deviation σ, and the second threshold value is set to about 3 times (3σ) of the standard deviation. The

  On the other hand, in this embodiment, as described above, the background matching degree is calculated as the hit rate β that is the index value for each target block. Therefore, by using the index value β of the degree of background coincidence, it is possible to perform addition at a larger addition rate in the background still image portion.

  61 shows the addition rate conversion output characteristics of the addition rate conversion output units 608 and 609 when the hit rate β is considered.

  That is, in the example of FIG. 61, the first threshold value for the difference value is changed according to the hit rate β. That is, when the hit rate β is the lowest and β = 0 (moving subject portion), the first threshold value is set to 1 times the standard deviation σ (1σ) as in the example of FIG. When the hit rate β is β> 0, the first threshold is changed to a larger value according to the value of the hit rate β.

  In the example of FIG. 61, the slope of the gradually decreasing straight line of the addition rate α from when the difference value becomes larger than the first threshold value is always constant. Therefore, in the example of FIG. 61, the second threshold value is automatically changed according to the change of the first threshold value.

  For example, as shown in FIG. 62, in the example of FIG. 60 in which the hit rate β is not considered (β = 0), the addition rate α is always 0.5 when the difference value is 2σ. On the other hand, when the hit rate is considered, as shown in FIG. 62, in the background still image portion where the hit rate β = 1.0, the addition rate α is α = 1, and the maximum addition rate is It will be added.

  In this way, in the background still image portion, a larger addition rate α is set according to the hit rate β, so that the expected noise reduction effect can be obtained. In addition, in the moving subject portion, the addition rate α can be kept low according to the hit rate β, so that the effect of reducing the double exposure state can be maintained.

  In the above example, only the first threshold value is changed according to the hit rate β, and the slope of the gradual decrease line of the addition rate α when the difference value becomes larger than the first threshold value is always constant. However, the second threshold value may also be changed according to the hit rate to change the slope of the gradual decrease line.

  Next, the contribution of the gain GA in the addition rate conversion output units 608 and 609 will be described.

  The gain GA is a parameter for further changing and controlling the addition rate α calculated as described above. As described above, the gain GA is set by the user or set according to the brightness of the image according to the shooting conditions. For example, when setting is performed by the user, the gain GA is increased in order to obtain a larger noise reduction effect by NR addition. Then, the addition rate α calculated as described above is controlled to be changed to a value corresponding to the gain GA.

  FIG. 63 shows the relationship among the gain GA, the hit rate β, and the addition rate α in the addition rate conversion output units 608 and 609. In FIG. 63, the addition rate α is shown as a value based on the reference addition rate αo with the addition rate when the hit rate β = 0 as the reference addition rate αo.

  In the example of FIG. 63, the gain GA can be changed and set in four stages of GA = 0, GA = 1, GA = 2, and GA = 3, depending on the user or shooting conditions. The hit rate β is also set in four stages: β = 0, β = 0.25, β = 0.5, and β = 1.0.

  As shown in FIG. 63, the addition rate α is obtained by multiplying the reference addition rate αo by a coefficient K (K ≧ 1), and the coefficient K depends on the combination of the gain GA and the hit rate β. It is a multiple.

  As described above, the addition rate conversion output unit 608 outputs the output addition rate αYA based on the pixel difference of the luminance component from the pixel difference absolute value ΔPxY, the noise standard deviation σY, the hit rate β, and the gain GA. To do. Also, the addition rate conversion output unit 608 outputs an output addition rate αYB based on the average difference of luminance components from the average difference absolute value ΔMeY, the noise standard deviation σY, the hit rate β, and the gain GA.

  For the color difference signal component, the average difference is not calculated, only the pixel difference absolute value is calculated, and the addition rate based on the color difference component is calculated in the same manner as the luminance component described above.

  FIG. 58 is a processing portion for the color difference component in the addition rate calculation unit 21. With reference to FIG. 58, the addition rate calculation process for the color difference component will be described.

  That is, blue color difference data Cbtg in the target image data TGv is supplied to the difference absolute value calculation unit 623 through the low-pass filter 621. Also, the blue color difference data Cbmc in the motion compensation image data MCv is supplied to the difference absolute value calculation unit 623 through the low-pass filter 622. The difference absolute value calculation unit 623 calculates the pixel difference absolute value ΔPxCb between the color difference data Cbtg and the color difference data Cbmc for each pixel, and supplies the calculated pixel difference absolute value ΔPxCb to the addition rate conversion output unit 625.

  Similarly to the case of the luminance component, a noise model memory 624 for the blue color difference component of the pixel value is provided. In the noise model memory 624, the noise standard deviation σ with respect to the blue color difference component value of the pixel value is It is calculated and stored in advance. As in the case of the luminance component, the noise model memory 624 stores a plurality of noise models corresponding to the illuminance Lx of the image.

  Then, the illuminance Lx for each target block and the color difference data Cbtg of the target image data TGv through the low-pass filter 621 are supplied to the noise model memory 624. Thus, the noise standard deviation σCb determined by the illuminance Lx and the color difference data Cbtg is obtained from the noise model memory 624 and supplied to the addition rate conversion output unit 625.

  Further, the addition rate conversion output unit 625 is supplied with a hit rate β from the motion vector calculation unit 15 and an external gain GA.

  The addition rate conversion output unit 625 has the same configuration as the addition rate conversion output units 608 and 609 described above. From this, the pixel difference absolute value ΔPxCb, the noise standard deviation σCb, the hit rate β, and the gain An addition rate αCb determined by the combination with GA is obtained.

  Similarly, red color difference data Crtg in the target image data TGv is supplied to the difference absolute value calculation unit 633 through the low-pass filter 631. The red color difference data Crmc in the motion compensation image data MCv is supplied to the difference absolute value calculation unit 633 through the low-pass filter 632. The difference absolute value calculation unit 633 calculates the pixel difference absolute value ΔPxCr between the color difference data Crtg and the color difference data Crmc for each pixel, and supplies the calculated pixel difference absolute value ΔPxCr to the addition rate conversion output unit 635.

  Similarly to the case of the luminance component, a noise model memory 634 for the red color difference component of the pixel value is provided, and the noise standard deviation σ with respect to the red color difference component value of the pixel value is stored in the noise model memory 634. It is calculated and stored in advance. As in the case of the luminance component, the noise model memory 634 stores a plurality of noise models corresponding to the illuminance Lx of the image.

  The illuminance Lx for each target block and the color difference data Crtg of the target image data TGv through the low-pass filter 631 are supplied to the noise model memory 634. As a result, the noise standard deviation σCr determined by the illuminance Lx and the color difference data Crtg is obtained from the noise model memory 634 and supplied to the addition rate conversion output unit 635.

  Further, the addition rate conversion output unit 635 is supplied with a hit rate β from the motion vector calculation unit 15 and an external gain GA.

  The addition rate conversion output unit 635 has the same configuration as the addition rate conversion output units 608 and 609 described above. From this, the pixel difference absolute value ΔPxCr, the noise standard deviation σCr, the hit rate β, and the gain An addition rate αCr determined by the combination with GA is obtained.

  In this embodiment, the addition rate is also obtained based on the pixel dispersion value of the target image. The variance value is obtained for the target pixel (own pixel) and, for example, eight pixels in the vicinity thereof. This variance value is obtained for each of the luminance component and the color difference component. Then, an addition rate is calculated from all of the obtained dispersion values.

  That is, as shown in FIG. 57, the luminance data Ytg of the image data TGv of the target block is supplied to the spatial dispersion filter 610 through the low-pass filter 601. Then, from this spatial dispersion filter 610, a dispersion value VaY is obtained for a total of nine pixels, that is, the target pixel (self pixel) and, for example, eight pixels in the vicinity thereof.

  As shown in FIG. 58, the blue color difference data Cbtg of the image data TGv of the target block is supplied to the spatial dispersion filter 626 through the low-pass filter 621. Then, from this spatial dispersion filter 626, a dispersion value VaCb is obtained for a total of nine pixels, for example, the pixel of interest (own pixel) and, for example, eight pixels in the vicinity thereof.

  Further, the red color difference data Crtg of the target block image data TGv is supplied to the spatial dispersion filter 636 through the low-pass filter 631. Then, from this spatial dispersion filter 636, a dispersion value VaCr is obtained for a total of nine pixels, for example, the pixel of interest (own pixel) and, for example, eight pixels in the vicinity thereof.

  Then, the variance values VaY, VaCb, and VaCr are added by the adding unit 611 and then supplied to the addition rate conversion output unit 612, as shown in FIG. Further, the addition rate conversion output unit 612 is supplied with the hit rate β from the motion vector calculation unit 15 and the gain GA from the outside.

  The addition rate αC obtained by the addition rate conversion output unit 612 is intended to suppress the addition rate when the image of the target block has few texture components and the luminance level changes gently (in the case of gradation). In the image processing divided into blocks as in this embodiment, it is considered that the block boundary becomes very conspicuous due to human visual effects, especially when the luminance level change is gentle.

  In order to realize the above object, basically, the lower the dispersion value in the block, the lower the addition rate. In the addition rate conversion output unit 612 of this embodiment, as shown in FIG. 64 (A), the horizontal axis represents the variance value, and the variance value is divided into several steps for each predetermined step width, and the output addition rate. Is determined according to the variance value of each stage.

  In the example of FIG. 64 (A), the variance value is divided into five steps for each step width 8, the addition rate for the variance value 0-7 is 0.0, the addition rate for the variance value 8-15 is 0.25, and the variance is The addition rate for the value 16-23 is 0.5, the addition rate for the variance value 24-31 is 0.75, and the addition rate for the variance value 32 or more is 1.0.

  By the way, when images are overlapped, the boundary for each block as described above is conspicuous because there is a difference in the images to be overlapped. In other words, with respect to the background still image portion, there is no difference in the superimposed images, so there is no need to suppress the addition rate according to the variance value.

  Therefore, in this embodiment, the addition rate conversion output unit 612 controls the step width of the variance value according to the hit rate β, and the step width of the variance value is reduced as the hit rate β is increased. As a result, for the background still image portion, the reduction of the addition rate according to the variance value is suppressed, and the addition can be performed at the highest possible addition rate.

  That is, in the example of FIG. 64, when the hit rate β is a low value such as β = 0, for example, the step width of the variance value is “8” as shown in FIG. When the hit rate β is a value such as β = 0.5, for example, the step width of the variance value is “4” as shown in FIG.

  Furthermore, when the hit rate β is β = 1 indicating that it is a background still image portion, the step width of the variance value is set to “0” as shown in FIG. That is, at this time, the addition rate is not reduced by the variance value.

  As described above, the gain GA is also supplied to the addition rate conversion output unit 612, and the output addition rate can be controlled by the gain GA. In the case of this example, the step width of the variance value is changed and controlled by the gain GA in combination with the hit rate β.

  FIG. 65 shows the relationship among the gain GA, the hit rate β, and the step width in the addition rate conversion output unit 612. In the example of FIG. 65 as well, the gain GA can be changed and set in four stages of GA = 0, GA = 1, GA = 2, and GA = 3 by the user or the shooting conditions. The hit rate β is also set in four stages: β = 0, β = 0.25, β = 0.5, and β = 1.0.

  In the example of FIG. 65, the step width is configured to be realized by a bit shift operation.

  As shown in FIG. 64, in this embodiment, when calculating the addition rate using the variance value, a plurality of steps using a step width is set for the variance value. Instead of this, a configuration may be adopted in which linear interpolation is performed between the plurality of stages. Further, instead of linear interpolation, quadratic curve interpolation or cubic curve interpolation may be performed.

  As described above, each of the addition rate conversion output units 608, 609, 612, 625, and 635 responds to the luminance difference value, the color difference difference value, the luminance average difference value, the variance value, and the like, and the value corresponding to the hit rate β. As a result, addition rates αYA, αYB, αC, αCb, and αCr are calculated.

  In this embodiment, the addition rates αYA, αYB, αC, αCb, and αCr are combined to obtain the output addition rate α of the addition rate calculation unit 21. That is, as shown in FIG. 57, the addition rates αYA and αYB from the addition rate conversion output units 608 and 609 are multiplied by the multiplier 613, and the multiplication result is supplied to the multiplier 614. The multiplier 614 is supplied with the addition rate αCb from the addition rate conversion output unit 625 and is multiplied by the multiplication result of the multiplier 613.

  Then, the multiplication result of the multiplier 614 is supplied to the multiplier 615. The multiplier 615 is supplied with the addition rate αCr from the addition rate conversion output unit 635 and is multiplied by the multiplication result of the multiplier 614. The multiplication result of the multiplier 615 is supplied to the multiplier 616. The multiplier 616 is supplied with the addition rate αC from the addition rate conversion output unit 612 and is multiplied by the multiplication result of the multiplier 615.

  As described above, the multiplier 616 obtains a multiplication result obtained by multiplying all the calculated addition rates αYA, αYB, αC, αCb, αCr, and outputs the result as the output addition rate α of the addition rate calculation unit 21. .

  In this way, the addition rate calculation unit 21 calculates the addition rate α as a value corresponding to the hit rate β in units of pixels and supplies it to the addition unit 17.

  In the addition unit 17, the image data TGv of the target block and the image data MCv of the motion compensation block are added in units of pixels according to the addition rate α for each pixel from the addition rate calculation unit 21. The added image data from the adding unit 17 is recorded in the recording / reproducing device unit 5 via the image memory unit 4 and the still image codec unit 18.

  According to the embodiment described above, the addition rate α can be increased in the still image portion by using the hit rate β that is the index value of the degree of background matching for each target block. Produces a remarkable effect that a large image can be obtained.

  For example, consider an NR image for a target image with noise as shown in FIG. 66 (FIG. 66A) and a reference image (FIG. 66B). In FIG. 66, a portion indicated by a circle is a moving subject portion between the target image and the reference image. The two images in this example are all background still image portions except for the moving subject portion.

  66. Block matching is performed on the two images in the example of FIG. 66, the local motion vector LMV is obtained for each target block, and the local motion vector LMV is illustrated as shown in FIG.

  Then, without considering the hit rate β as described above, using this local motion vector LMV, the motion compensation image is performed from the reference image, and the addition rate distribution image as the result of the addition rate determination is obtained as a monochrome image. As shown in FIG. In FIG. 68, as the addition rate increases, it is expressed in white, and as the addition rate decreases, it is expressed in black.

  In FIG. 67, the circled portion is the background still image portion, and the correct local motion vector LMV is obtained. However, since the hit rate β is not considered, the corresponding portion in FIG. Indicates that the addition rate is low. That is, it can be seen that the expected noise reduction effect cannot be obtained.

  Next, as described in the above-described embodiment, the global motion GM is obtained for the two images shown in FIG. 66, and the global motion vector GMV generated from the global motion GM is illustrated in FIG. 69 (A). It will be as shown in

  Then, the degree of coincidence between the generated global motion vector GMV and the local motion vector LMV is checked. As a result, the motion vector to be output to the motion compensation image generation unit 16 is illustrated in FIG. 69 (B). As shown. The large motion vector for the lower right image portion in FIG. 69B is the moving subject image portion shown in FIG. As can be seen from FIG. 69, according to this embodiment, the background still image portion and the moving subject portion are correctly distinguished and determined.

  Then, when the hit rate β calculated when generating the motion vector of FIG. 69B is shown in correspondence with the image, it becomes as shown in FIG. However, in FIG. 70, the degree of coincidence between the local motion vector LMV and the global motion vector GMV is determined by binary values of coincidence and non-coincidence, and the hit rate β is a binary value of β = 0 and β = 1. It shows as the case obtained by. The white portion is a portion where β = 1, and the black portion is a portion where β = 0.

  FIG. 71 shows the addition rate α obtained by this embodiment corresponding to the image. In FIG. 71, with respect to the addition rate α, 0.0 to 1.0 is divided into 128 gradations, and the addition rate is displayed so as to decrease from white to black (closer to α = 0.0).

  Comparing FIG. 67 and FIG. 71, according to this embodiment, it can be seen that the addition rate α is increased in the background still image portion. Thereby, according to this embodiment, the expected noise reduction effect is acquired.

  The addition rate conversion output units 608, 609, 612, 625, and 635 include conversion tables that output the output addition rates αYA, αYB, αC, αCb, and αCr for each of the four parameter inputs described above. A ROM configuration can be adopted.

[Third Embodiment]
In the second embodiment described above, the hit rate β is calculated in units of target blocks. Then, the addition rate α is controlled using the hit rate β of the target block unit. For this reason, although the addition rate α is calculated in units of pixels, there is a problem that the boundary of the block unit of the NR image is easily noticeable.

  Therefore, in the third embodiment, the hit rate β calculated for each target block is interpolated with reference to the hit rate of the target block adjacent to the own block so that the hit rate β changes spatially and smoothly. To be processed.

  According to the third embodiment, the block boundary Bd in the above-described process 1 is also inconspicuous, and therefore, the block boundary in the process 2 and later is further inconspicuous.

<First Example of Hit Rate β Interpolation Processing>
In the first example, the hit rate β is interpolated in units of target blocks. 72 and 73 show an example of interpolation processing in the case of the first example. 72 and 73, Bk indicates each of the target blocks, and the numerical value described in the target block Bk indicates the value of the hit rate β. Here, for simplicity of explanation, the calculated hit rate β is a binary value of β = 1.0 and β = 0.5.

  The example in FIG. 72 refers to the hit rate β of another target block Bk adjacent to the own target block Bk, and interpolates in a direction to reduce the hit rate of the own target block Bk so as to change spatially smoothly. This is an example. That is, when the hit rate β calculated for each target block Bk is as shown in FIG. 72A, the hit rate β of a predetermined target block Bk is obtained by interpolation as shown in FIG. 72B. Is changed.

  In the example of FIG. 72, when the hit rate β of its own target block Bk is β = 1.0 and the hit rate β of another adjacent target block Bk is β = 0.0, The target block Bk is set as a target block to be interpolated. Then, the hit rate β of the target block to be interpolated is changed from β = 1.0 to β = 0.5 in this example. As a result, the hit rate β of the target block Bk unit is spatially smoothly changed as shown in FIG. 72 (B).

  The example in FIG. 73 refers to the hit rate β of another target block Bk adjacent to the own target block Bk, and interpolates in a direction to increase the hit rate of the own target block Bk so as to change spatially smoothly. This is an example. That is, when the hit rate β calculated for each target block Bk is as shown in FIG. 73A, the hit rate β of a predetermined target block Bk is obtained by interpolation as shown in FIG. 73B. Is changed.

  In the example of FIG. 73, when the hit rate β of the target block is β = 0.0 and the hit rate β of another adjacent target block is β = 1.0, the target The block Bk is set as a target block to be interpolated. Then, the hit rate β of the target block to be interpolated is changed from β = 0.0 to β = 0.5. As a result, the hit rate β of the target block Bk unit is spatially smoothly changed as shown in FIG. 73 (B).

  Thereby, even if the hit rate β is calculated for each target block, it is possible to make the boundary of the block unit inconspicuous.

<Second example of interpolation processing of hit rate β>
In the second example, the hit rate β is interpolated in units of fine blocks obtained by further dividing the target block Bk. 74 and 75 show an example of interpolation processing in the case of this second example. 74 and 75, Bk represents each of the target blocks, SBk represents a fine block obtained by dividing the target block Bk into four, and SSBk represents a fine block obtained by dividing the target block Bk into eight. The numerical values described in the target block Bk and the fine block indicate the value of the hit rate β. Also in this second example, for simplicity of explanation, the calculated hit rate (hit rate before interpolation processing) β is a binary value of β = 1.0 and β = 0.5.

  The example of FIG. 74 refers to the hit rate β of another target block Bk adjacent to its own target block Bk, and interpolates in a direction to reduce the hit rate of its own target block Bk so as to change smoothly spatially. This is an example.

  In the example shown in FIGS. 74A and 74B, interpolation is performed in units of fine blocks SBk obtained by dividing the target block Bk into four. That is, when the hit rate β calculated for each target block Bk is as shown in FIG. 74 (A), the fine block in the predetermined target block Bk is obtained by interpolation processing as shown in FIG. 74 (B). The hit rate β of SBk is changed.

  In the example of FIG. 74 (B), when the hit rate β of its own target block Bk is β = 1.0 and the hit rate β of another adjacent target block Bk is β = 0.0. In addition, the target block Bk is set as a target block to be interpolated. Then, the hit rate β of the target block to be interpolated is changed to a value of β = 1.0 or less in units of fine blocks SBk.

  In this case, instead of changing the hit rate β of all the fine blocks SBk in the own target block Bk to β = 0.5, it is adjacent to another target block Bk whose hit rate β is β = 0.0. The hit rate β of only the fine block SBk to be changed is changed to β = 0.5. Then, the hit rate of other fine blocks in the own target block Bk is held as β = 1.0. In the case of this second example, the adjacent target blocks Bk include those adjacent in the oblique direction. Therefore, the hit rate β of the fine block SBk in which the hit rate β of the target block Bk adjacent in the oblique direction is β = 0.0 is also changed to β = 0.5.

  In the example shown in FIGS. 74A and 74C, interpolation is performed in units of fine blocks SSBk obtained by dividing the target block Bk into eight. In this example, the hit rate β of the own target block Bk is β = 1.0, but the target block Bk to be interpolated is set to a hit rate β of 0.25, 0.5 for each fine block SSBk. , 0, 75, 1.0, etc., with a change width of 0.25.

  That is, when the hit rate β calculated for each target block Bk is as shown in FIG. 74 (A), the fine block in the predetermined target block Bk is obtained by interpolation processing as shown in FIG. 74 (C). The hit rate β of SSBk is changed.

  In the example of FIG. 74C, when the hit rate β of its own target block Bk is β = 1.0, and the hit rate β of another adjacent target block Bk is β = 0.0. In addition, the target block Bk is set as a target block to be interpolated. Then, the hit rate β of the target block Bk to be interpolated is changed in units of fine blocks SSBk.

  In this case, the hit rate β of the fine block SSBk adjacent to another target block Bk whose hit rate β is β = 0.0 is changed to β = 0.25. Then, the hit rate of the fine block adjacent to the fine block SSBk changed to β = 0.25 in the own target block Bk is assumed to be β = 0.5.

  Then, in the target block Bk to be interpolated, the hit rate β of the fine block SSBk adjacent to another target block Bk whose hit rate β is β = 1.0 remains β = 1.0. Then, the hit rate of the fine block adjacent to the fine block SSBk in which β = 1.0 in the target block Bk remains β = 0.75. Then, the hit rate of the fine block adjacent to the fine block SSBk changed to β = 0.75 in the own target block Bk is assumed to be β = 0.5.

  Also in this example, the adjacent target blocks Bk include those adjacent in the oblique direction.

  As a result, the hit rate β for each target block Bk is spatially and smoothly changed as shown in FIG. 74 (B) or (C).

  Next, in the example of FIG. 75, the hit rate β of the target block Bk is increased so as to spatially smoothly change with reference to the hit rate β of another target block Bk adjacent to the target block Bk. This is an example of interpolation in the direction.

  The example shown in FIGS. 75A and 75B is a case where interpolation is performed in units of fine blocks SBk obtained by dividing the target block Bk into four. That is, when the hit rate β calculated for each target block Bk is as shown in FIG. 75 (A), by the interpolation process, as shown in FIG. 75 (B), the fine block in the predetermined target block Bk. The SBk hit rate β is changed.

  In the example of FIG. 75 (B), when the hit rate β of its own target block Bk is β = 0.0 and the hit rate β of another adjacent target block Bk is β = 1.0. In addition, the target block Bk is set as a target block to be interpolated. Then, the hit rate β of the target block to be interpolated is changed to a value of β = 0.0 or more for each fine block SBk.

  In this case, instead of changing the hit rate β of all the fine blocks SBk in the target block Bk to β = 0.5, the hit rate β is adjacent to another target block Bk with β = 1.0. The hit rate β of only the fine block SBk to be changed is changed to β = 0.5. Then, the hit rate of other fine blocks in the own target block Bk is held as β = 0.0. In the case of this second example, the adjacent target blocks Bk include those adjacent in the oblique direction. Therefore, the hit rate β of the fine block SBk in which the hit rate β of the target block Bk adjacent in the oblique direction is β = 0.0 is also changed to β = 0.5.

  In the example shown in FIGS. 75A and 75C, interpolation is performed in units of fine blocks SSBk obtained by dividing the target block Bk into eight. In this example, the hit rate β of the own target block Bk is β = 0.0. In the target block Bk to be interpolated, the hit rate β is set to 0.25, 0.5 for each fine block SSBk. , 0, 75, 1.0, etc., with a change width of 0.25.

  That is, when the hit rate β calculated for each target block Bk is as shown in FIG. 75A, the fine block in the predetermined target block Bk is obtained by interpolation processing as shown in FIG. 75C. The hit rate β of SSBk is changed.

  In the example of FIG. 75C, when the hit rate β of its own target block Bk is β = 0.0, and the hit rate β of another adjacent target block Bk is β = 1.0. In addition, the target block Bk is set as a target block to be interpolated. Then, the hit rate β of the target block Bk to be interpolated is changed in units of fine blocks SSBk.

  In this case, the hit rate β of the fine block SSBk adjacent to another target block Bk whose hit rate β is β = 1.0 is changed to β = 0.75. Then, the hit rate of the fine block adjacent to the fine block SSBk changed to β = 0.75 in the own target block Bk is assumed to be β = 0.5.

  Then, in the target block Bk to be interpolated, the hit rate β of the fine block SSBk adjacent to another target block Bk whose hit rate β is β = 0.0 remains β = 0.0. Then, the hit rate of the fine block adjacent to the fine block SSBk in which β = 0.0 in the own target block Bk is set to β = 0.25. Then, the hit rate of the fine block adjacent to the fine block SSBk changed to β = 0.25 in the own target block Bk is assumed to be β = 0.5.

  Also in this example, the adjacent target blocks Bk include those adjacent in the oblique direction.

  As a result, the hit rate β of the target block Bk unit is spatially smoothly changed as shown in FIG. 75 (B) or (C).

[Fourth Embodiment]
In the above-described embodiment, the local motion vector calculation unit 154 performs hierarchical block matching to calculate the local motion vector LMV on the base surface. Then, the motion vector reliability index value calculation unit 156 calculates the reliability index Ft of the local motion vector LMV on the base surface.

  Further, the motion vector calculation unit 157 generates a reliability index value Ft for the basal plane local motion vector LMV, and extracts a highly reliable local motion vector LMV using the reliability index value Ft. I made it. Then, the global motion GM is calculated using the highly reliable local motion vector LMV, and the global motion vector GMV is calculated from the global motion GM.

  By the way, the basal plane basal plane motion vector MVb can be obtained by multiplying the reduced plane motion vector MVs and the intermediate plane motion vector MVm by the inverse of the image reduction magnification with respect to the basal plane. Therefore, in order to calculate the global motion GM, the global motion GM can be obtained from the reduced surface motion vector MVs and the intermediate surface motion vector MVm without obtaining the base surface motion vector MVb.

  For example, when the global motion GM is obtained from the reduced plane motion vector MVs, first, the local motion vector calculation unit 154 calculates the local motion vector LMV (reduced plane motion vector MVs) on the reduced plane.

  Then, the motion vector reliability index value calculation unit 156 calculates a reliability index Ft for the calculated reduced plane motion vector MVs. Also, the motion vector calculation unit 157 extracts a highly reliable local motion vector LMV using the reliability index value Ft for the local motion vector LMV of the reduced surface. Then, the motion vector calculation unit 157 calculates the global motion GM and the global motion vector GMV using the highly reliable local motion vector LMV.

  It can be calculated in the same manner when using the global motion GM from the intermediate plane motion vector MVm obtained by performing block matching in the hierarchy up to the intermediate plane. However, the offset value OFS and the gain g when correcting the SAD value at the position of the global motion vector GMV are appropriately set according to each layer.

  The merit of obtaining the global motion GM and the global motion vector GMV using the local motion vectors of the reduced surface and the intermediate surface in this way is as follows.

  First, as described above, since the low-pass filter is applied to the reduction surface and the intermediate surface, noise is removed, and as a result, the obtained local motion vector is not easily affected by noise. It is.

  The other is that the number of target blocks is reduced on the reduction plane and the intermediate plane, so the number of local motion vectors is reduced, the calculation cost is reduced, and the processing time and time are reduced, resulting in faster processing. Is possible.

  As in the example described above, the matching processing block units for the reduced surface, the intermediate surface, and the base surface are usually the same size due to hardware limitations. For this reason, the number of target blocks (= the number of local motion vectors) on the reduced surface with a small image size is relatively smaller than when block matching is performed only on the base surface.

  When the global motion and the global motion vector GMV are obtained from the reduced plane motion vector, the motion vector detection process on the intermediate plane and the base plane can be omitted. Further, when the global motion and the global motion vector GMV are obtained from the intermediate plane motion vector, the motion vector detection process on the base plane can be omitted. Therefore, the processing speed can be increased.

  In particular, when the global motion GM and the global motion vector GMV are obtained using the reduced plane motion vector MVs, the merit is great.

  FIG. 76 shows an overall flow of processing when the global motion and the global motion vector GMV are obtained using the reduced surface motion vector MVs and compared with the local motion vector LMV on the base surface. FIG. 76 shows the processing for one set of target frame and reference frame.

  First, the control unit 155 of the motion vector calculation unit 15 determines the division start point for the frame image of the reduced surface to be processed as described in the first embodiment, and the block division position is the previous frame. It is determined so as not to overlap with the block division position at (step S501).

  Next, the motion vector calculation unit 15 performs reduction plane block matching to obtain local motion vectors LMVs for each target block on the reduction plane. Then, the reliability of the reduced plane local motion vector LMVs for each target block is evaluated, and the global motion GM is calculated using only the reduced plane local motion vector LMVs with high reliability. Then, a global motion vector GMV is calculated from the calculated global motion (step S502).

  Next, for the hierarchized block matching, reduced surface re-block matching is performed, and the reduced surface local motion vector LMVs is recalculated (step S503).

  Next, a search range for a reference block with respect to the target block on the intermediate plane is obtained from the reduced plane local motion vector LMVs, and intermediate plane block matching is performed. Then, the intermediate plane local motion vector LMVm is calculated (step S504).

  Next, a search range for a reference block with respect to the target block on the base plane is obtained from the intermediate plane local motion vector LMVm, and base plane block matching is performed. Then, a basal plane local motion vector LMVb is calculated. Then, using the basal plane local motion vector LMVb and the global motion vector GMV calculated in step S501, the above-described background matching degree evaluation determination is performed for each target block. Based on the evaluation determination, the hit rate β is obtained, and the NR processing motion vector MVnr is detected (step S505).

  Next, the motion compensation image generation unit 16 generates a motion compensation image in block units from the image data of the reference frame using the NR processing motion vector MVnr (step S506).

  Further, the addition rate calculation unit 21 calculates the addition rate α in units of pixels as described above based on the hit rate β calculated in step S505 (step S507). Then, the adding unit 17 adds the motion compensation image generated in step S506 and the target image in units of blocks at the addition rate α calculated in step S507 to generate an NR image (step S508).

  In the description of the processing flow of FIG. 76 described above, the global motion GM and the global motion vector GMV obtained at the time of motion vector detection by the first block matching are based on the base motion vector of the hierarchical re-block matching. Only applied.

  However, even in the reduced and intermediate planes of the hierarchical re-block matching, the target block background matching degree evaluation using the global motion vector GMV, the reduced plane local motion vector LMVs, and the intermediate plane local motion vector LMVm may be performed. Good. In this case, the resulting motion vector MVnr is used as a motion vector output for each target block on the reduced plane and the intermediate plane.

  In such a case, it is possible to acquire the motion vector (local motion vector) for each target block on the reduced surface and the intermediate surface with higher accuracy than usual. The same applies to the basal plane. At that time, as described above, the offset value OFS and the gain g are appropriately set according to each layer.

  FIGS. 77A to 77D show examples of block matching results and detected motion vectors when processing is performed according to the processing flow of FIG. The example of FIG. 77 shows a processing result when the above-described processing is applied to the above-described target image of FIG. 78A and the reference image of FIG. 78B.

  FIG. 77A shows the detection result of the reduced surface local motion vector LMVs when reduced surface block matching is performed on the target image of FIG. 78A and the reference image of FIG. 78B. ing.

  FIG. 77 (B) shows the reduced-plane local motion vector LMVs and its target block, which are evaluated to be reliable for the reduced-plane local motion vector LMVs, and have high reliability. In FIG. 77 (B), the block in the lower right corner is a moving subject portion, and has a reduced reduction plane local motion vector LMVs.

  FIG. 77C shows the global motion vector GMV for each target block calculated based on the global motion GM calculated from the base surface local motion vector LMVb. Further, FIG. 77 (D) shows a case where the motion vector is detected again and the background portion and the moving subject portion are evaluated as described above to detect the NR processing motion vector MVnr.

  From FIG. 77 (D), it can be seen that the motion vectors of the background portion are neatly arranged as the global motion vector GMV.

  In FIG. 77 (D), the large lower right motion vector is assumed not to output the global motion vector GMV as the NR processing motion vector in the evaluation determination using the local motion vector LMV and the global motion vector GMV. This is a moving subject block. A white vector is a block in which there is no reference vector that matches the global motion vector GMV in the effective pixel region within the search range.

  FIG. 77 is created to show the effect in an easy-to-understand manner, and actually has a smaller target block size. As described above, it is more effective to generate the hit rate β in units of small blocks.

  However, in the case of an application that detects and tracks a moving subject region and a vector in an image in a coarse unit, a motion based on an evaluation using the global motion vector GMV in a coarse unit as shown in FIG. 77 (D). It is preferable to output the vector MVnr. At that time, if information on all blocks in the screen is not necessary, and if reliability of moving subject detection is emphasized, the reliability at the time of the first motion vector detection is high as shown in FIG. 77 (B). The moving subject block may be separated only from the vector.

[High accuracy by interpolation processing]
Incidentally, it is necessary to consider that the accuracy of the reduced surface motion vector and the intermediate surface motion vector is low because the image of the base surface is reduced on the reduced surface and the intermediate surface.

  Therefore, in order to improve this low-accuracy problem, interpolation processing is performed when a reduced plane motion vector or an intermediate plane motion vector is used. That is, the SAD value of the reduced surface reference block or the intermediate surface reference block in the vicinity of the reduced surface reference block position indicated by the calculated reduced surface motion vector or the intermediate surface motion vector or the intermediate surface reference block position and the position information thereof are used. To perform interpolation processing. By this interpolation processing, it is possible to detect a reduced plane motion vector or an intermediate plane motion vector with pixel accuracy. This interpolation process will be described below by taking the case of a reduction plane as an example.

  For example, when block matching is performed on a reduced surface that has been reduced to ¼ both vertically and horizontally, the reduced surface motion vector is a 4-pixel precision motion vector. However, in the base plane reference frame, it is clear that the base plane motion vector MVb with 1 pixel accuracy exists in the vicinity of the motion vector obtained by multiplying the reduced plane motion vector MVs by n.

  Therefore, as shown in FIG. 79, when the minimum SAD value 601 on the reduction plane is obtained, interpolation is performed using a plurality of neighboring SAD values 602, 603, 604, 605, for example, four adjacent in the vertical and horizontal directions. It can be considered to perform processing and detect a motion vector with pixel accuracy. In this case, the interpolation magnification is 4 times.

  For example, it is considered to calculate a pixel-precision motion vector from a reduced surface SAD table that has been subjected to matching processing in units of n pixels, for example, by interpolating the SAD table using a quadratic curve. In this case, instead of quadratic curve approximation interpolation, linear interpolation or higher-order approximation curve interpolation of cubic or higher may be used. However, in this example, a quadratic curve is used in consideration of the balance between accuracy and hardware implementation. Approximate interpolation is used.

  In this quadratic curve approximate interpolation, as shown in FIG. 79, the minimum value Smin of the SAD value of the reduced surface SAD table indicated by the reduced surface motion vector of n pixel accuracy and a plurality of positions near the position of the minimum value Smin. SAD values (referred to as neighborhood reduced surface SAD values) are used. In this example, as the neighboring reduced surface SAD values, four neighboring SAD values Sx1, Sx2, and Sy1, which are adjacent in the X direction (horizontal direction) and the Y direction (vertical direction) at the position of the minimum value Smin on the reduced surface, Use Sy2.

  First, in the X direction (horizontal direction) interpolation, as shown in FIG. 80, the minimum value Smin of the reduced surface SAD value and the adjacent reduced surface SAD values Sx1 and Sx2 of the two neighboring points in the X direction (horizontal direction). On the other hand, a quadratic approximate curve 700 is applied. That is, a quadratic approximate curve 700 is obtained that passes through the minimum value Smin and two adjacent reduced surface SAD values Sx1 and Sx2 in the vicinity of the X direction (horizontal direction). Then, as shown in FIG. 80, the coordinate at which the quadratic curve 700 takes the minimum is the X coordinate Vx of the reduced surface motion vector (high-precision reduced surface motion vector) that becomes the minimum value SXmin of the SAD value with pixel accuracy. . The equation of quadratic curve approximate interpolation at this time is shown in the following equation (Equation e).

SXmin = 1/2 * (Sx2-Sx1) / (Sx2-2Smin + Sx1) (Expression e)
The X coordinate Vx which the minimum value SXmin of the SAD value with the pixel accuracy obtained by this calculation formula (formula e) takes on the SAD table becomes the X coordinate Vx which becomes the minimum value of the reduced surface SAD value with the pixel accuracy.

  The division of this calculation formula (formula e) can be realized by a plurality of subtractions. If the pixel accuracy to be obtained is, for example, an accuracy of 1/4 of the pixel pitch on the reduction plane, it can be obtained by subtraction twice, so both the circuit scale and the computation time are small, and a quadratic approximate curve Performance that is almost the same as cubic curve interpolation, which is considerably more complicated than interpolation, can be realized.

  Similarly, in the interpolation in the Y direction (vertical direction), a quadratic order is obtained with respect to the minimum value Smin of the reduction surface SAD value and the two adjacent reduction surface SAD values Sy1 and Sy2 in the vicinity of the Y direction (vertical direction). Fit an approximate curve. Then, the Y coordinate that takes the minimum value SYmin of this quadratic approximate curve becomes the Y coordinate Vy that is the minimum value of the SAD value with pixel accuracy. The equation of quadratic curve approximate interpolation at this time is shown in the following equation (expression f).

SYmin = 1/2 × (Sy2−Sy1) / (Sy2-2Smin + Sy1) (Formula f)
As described above, by performing approximation of the quadratic curve twice in the X direction and the Y direction, a highly accurate reduced surface motion vector (Vx, Vy) with pixel accuracy can be obtained.

  In the above description, the minimum value of the reduction surface SAD value and the two reduction surface SAD values in the vicinity in the X direction (horizontal direction) and the Y direction (vertical direction) are used, but the reduction surface SAD in the vicinity in each direction. The value may be two or more points. Further, instead of the quadratic curve in the X direction and the Y direction, for example, an approximate curve may be fitted in an oblique direction. Furthermore, an approximate curve may be applied by adding an oblique direction to the X direction and the Y direction.

  FIG. 81 shows that a vector detection result with pixel accuracy can be obtained from the value of the SAD table with accuracy of n pixels by using the above means and procedure. The horizontal axis in FIG. 81 is the interpolation magnification and represents how many times the resolution is increased in the one-dimensional direction. Since the SAD table is two-dimensional, the table area is reduced at a ratio of the square, whereas the error due to interpolation increases only to a linear level, so that the usefulness of the above-described interpolation method can be seen.

  Needless to say, the interpolation method described with reference to FIGS. 79 to 81 is applicable not only to the reduction plane but also to the intermediate plane and the base plane. When the above interpolation method is applied to the base plane, a local motion vector LMV with an accuracy (subpixel accuracy) smaller than the pixel pitch is obtained.

  When the local motion vector LMV with sub-pixel accuracy is obtained on the base plane, the global motion vector GMV output as the motion vector MVnr for NR processing outputs the local motion vector LMV with sub-pixel accuracy on the base plane that matches this. Good. That is, for the block determined to be the background, the global motion vector GMV calculated from the global motion GM may be output, or a local motion vector LMV that matches the global motion vector GMV may be output. In either case, it is better to output a highly accurate local motion vector LMV as the global motion vector GMV.

  In addition, when the local motion vector LMV with sub-pixel accuracy is calculated on the base plane, there is also an advantage that the motion compensation image generation unit 16 can also generate an interpolation image with sub-pixel accuracy as a motion compensation image.

[Effect of the above-described embodiment]
As described above, according to the above embodiment, motion vectors for all target blocks in an image can be detected with high accuracy. Furthermore, it is possible to determine whether the image is a background image portion or a moving subject portion, and a hit rate that is an index value can be obtained. The image addition rate can be increased according to the hit rate.

  Therefore, it is possible to detect a local motion vector with high accuracy from an image with a high noise level, which has been difficult in the past, and to extract a background still image region and a moving subject region in the entire image. In, images can be superimposed at a high addition rate. For this reason, an NR image that can achieve the desired noise reduction effect can be obtained.

  Since the block division position is offset-changed for each frame as described above, there is an effect that the block boundary for the superimposed image after addition becomes inconspicuous.

[Other Embodiments and Modifications]
In the above description, the background / moving subject determination unit 150 performs the evaluation determination of the degree of background matching for all target blocks in the target frame.

  However, in this embodiment, the global motion is obtained by performing a convergence calculation using only the highly reliable local motion vector LMV. Among the highly reliable local motion vectors LMV, the local motion vector LMV excluded in the global motion convergence calculation process in FIGS. 44 and 45 is likely to be the local motion vector LMV of the moving subject block. Further, in the global motion convergence calculation, the local motion vector LMV finally used for the global motion calculation can be determined to be the background block.

  Therefore, when calculating global motion, it is not necessary to make another evaluation determination as to whether the block exhibiting the evaluated local motion vector LMV with high reliability is the background or the moving subject. Therefore, re-block matching and background / moving subject determination processing can be omitted for the target blocks of the local motion vector LMV with high reliability. As a result, the processing time and the calculation amount can be reduced.

  In the global motion convergence calculation, since the local motion vector LMV finally used for calculating the global motion can be determined to be a background block, the hit rate β can be set to β = 1.0. it can.

  When the hit rate β is 1.0, the addition rate α between the target block and the compensation image block is set to 1.0, so that the background portion is positively and strongly added, resulting in a large NR effect. I can expect.

  In the above-described embodiment, the motion compensation image is motion compensated in units of target blocks. However, since global motion GM is obtained, motion compensation can be performed in units of pixels. That is, for example, for a target block image with a high hit rate β such as 1.0 or 0.75 and a high degree of background matching, a global motion vector is calculated from the global motion for each pixel. Then, motion compensation is performed using the calculated global motion vector for each pixel. Since the motion compensation result may not match the pixel position, interpolation processing is performed to generate a motion compensation pixel, thereby generating a motion compensation image.

  Thus, if motion compensation is performed in units of pixels, it can be expected that a smoother motion compensated image can be obtained.

  In the above-described embodiment, the SAD value is detected as the correlation value, but it goes without saying that the correlation value is not limited to the SAD value.

  In the above-described embodiment, the motion detection process and the image superimposition process for still images are performed by capturing a plurality of captured images into the image memory unit 4 and regarding the captured plurality of captured images as described above. Smooth motion detection and image overlay processing. However, it is also possible to process a plurality of images to be processed in real time as in the case of moving image shooting.

  Needless to say, the image information to be detected by the motion vector is not limited to the captured image information.

  In the above-described embodiment, the reliability of the motion vector is determined based on the motion vector reliability index value. However, not only the difference or ratio between the first maximum value of the correlation value and the second maximum value of the correlation value, but also the reference vector that takes the first maximum value of the correlation value and the reference vector that takes the second maximum value of the correlation value The reliability may be determined by further using the positional difference between and. Further, the third maximum value of the correlation value, further, the maximum value of the higher rank order and the position distribution of the reference vector taking the maximum value are further referred to to determine the reliability. You may make it do.

  In the above-described embodiment, the weighting factor W corresponding to the reliability index value of the local motion vector LMV is binarized into 0 and 1 in order to reduce the calculation load in the global motion calculation process. It was used. However, of course, the global motion calculation process may be performed using the weighting factor W of, for example, 0 or more and 1 or less obtained by normalizing the reliability index value of the local motion vector LMV as it is.

  Further, the global motion is not calculated by signal processing using affine transformation as in the above-described embodiment, but is detected as a displacement applied to the entire image sensor using, for example, an acceleration sensor such as a gyro. You can also. Then, using the detected global motion, the global motion vector GMV is obtained in the same manner as described above, and the obtained global motion vector GMV and the local motion vector LMV are compared to obtain the degree of coincidence between them. Can do.

  In this case, as the global motion vector GMV, only one vector obtained from the global motion may be compared with each local motion vector LMV. Further, it may be determined that the global motion has occurred at the center position of the image (frame), and the global motion vector GMV of each target block may be obtained from the global motion using the center position as the origin coordinate.

  In addition, the global motion vector GMV for each target block for the image to be superimposed can be generated outside. In that case, the global motion vector GMV does not need to be calculated as described above in the image processing apparatus, and may simply be acquired from the outside. Externally, the global motion vector GMV can be generated by calculation including only signal processing as described above or using a sensor such as a gyro and performing necessary calculations.

  In particular, for still images, images continuously shot in advance are stored, and when performing NR processing at a later time, global motion is calculated externally and global motion vector GMV is acquired. It is relatively easy.

  In the above-described embodiment, the present invention is applied to the case of motion vector detection and background / moving subject determination during still image shooting. However, the present invention performs motion vector detection and background / moving subject determination during moving image shooting. Of course, the present invention can also be applied.

  FIG. 82 shows an example of the hardware configuration of the imaging apparatus considering the time of moving image shooting. In FIG. 82, the same parts as those in FIG. 1 are denoted by the same reference numerals, and detailed description thereof is omitted.

  In the NR processing of the moving image in the example of FIG. 82, the captured image from the image sensor 11 that is input becomes the target image every 1V (V is a vertical period (frame or field)), and the output image after the addition Images before 1V and before 2V are reference images.

  The reason why the image before 2V as well as the image before 1V is referred to is that, in the case of an interlaced image, the matching rate of the stationary object with that before 2V is higher.

  Also in the system of FIG. 82, global motion is calculated from the local vector, hit rate information is generated depending on whether the motion vector of each block or each pixel in the target image matches the global motion, and the subsequent addition rate α is controlled. Thus, a moving image capturing system having a high NR effect is obtained.

  In the case of the example shown in FIG. 82, the codec unit 22 that performs MPEG compression coding is used.

  The above description has been for the case where the image information is the captured image information. However, it goes without saying that the image information targeted by the present invention is not limited to the captured image information.

  In the above-described embodiment, the addition rate is calculated in units of pixels. However, the addition rate may be calculated in units of target blocks.

  The above-described embodiment is a case where the present invention is applied to image addition processing by all NR processing, but it goes without saying that image processing for performing processing by dividing blocks is not limited to image addition processing.

  FIG. 83 shows a block diagram of an embodiment to which the present invention is applied when the image processing is image recognition processing.

  The embodiment of FIG. 83 includes a block division processing unit 1011 that divides input image information into blocks, an image recognition processing unit 1012 that performs image recognition processing on image information in units of blocks from the block division processing unit 1011, and a control unit 1010. It consists of.

  In the embodiment of FIG. 83, the control unit 1010 changes the block division position for the image information of the same frame to the block division processing unit 1011 and performs image recognition at the image recognition processing unit 1012 a plurality of times. Control to make it happen.

  A flowchart of the image recognition process in the embodiment of FIG. 83 is shown in FIG.

  First, the control unit 1010 designates the first block division start point for the input image information, and causes the block division processing unit 1011 to perform block division (step S601).

  The image recognition processing unit 1012 performs image recognition processing on the block divided into blocks (step S602). The image recognition process in step S602 may be performed for all blocks on one screen, or may be performed only for a block at a predetermined target position.

  When the image recognition processing unit 1012 completes the image recognition process for one frame at one block division position, in this example, the control unit 1010 performs image recognition for a predetermined number of times (a predetermined number of frames). It is determined whether or not processing has been performed (step S603). If it is determined that the image recognition has not been performed a predetermined number of times, the control unit 1010 returns to step S601 and instructs the block division processing unit 1011 to specify a block division start point for making a block division different from the previous one. Then, block division is performed.

  In step S602, the image recognition processing unit 1012 performs image recognition processing again on the re-divided block.

  If it is determined in step S603 that the image recognition processing has been repeated a predetermined number of times, the control unit 1010 ends the image recognition processing in the image recognition processing unit 1012. Then, the image recognition processing unit 1012 is instructed to output the best processing result among the predetermined number of times of image recognition processing (step S604). Then, this processing routine ends.

  In this embodiment, image recognition is performed on a block-by-block basis, but since the block position is a different position on the image, the image recognition result will be different. A good recognition result may be obtained at the division position.

  Therefore, according to the embodiment of FIG. 83, there is an effect that the image recognition rate can be improved.

  As described above, the image processing apparatus according to the present invention performs control so that the block division position is changed for each frame. In this case, the image information in the above-described embodiment is sequentially predetermined. The image is input at the frame rate. On the other hand, the image information of the image recognition target in the embodiment of FIG. 83 is the image information of the same one frame.

  In other words, according to the present invention, for image information of the same frame, when the block division position is changed for each image processing for one frame and image processing is performed, the image information input as a time-series signal is changed for each frame. Any of the cases of image processing can be applied.

  In the first to third embodiments described above, all image information for one frame (one screen) is divided into equal blocks. However, in image recognition or the like as in the embodiment of FIG. 83, it is also possible to divide a block so that a partial block at a specific position on a predetermined screen is processed.

  In the former case, since one screen is equally divided into blocks, the block division position is determined simply by determining the division start point of the image for one screen, so one division start point for one screen. Only need to be controlled. However, in the latter case, it is necessary to control the block division position by controlling each of the start points of each block.

  Note that the present invention is not limited to the above-described NR addition processing and image recognition processing, but can be applied to all image processing apparatuses and image processing methods that process image information by dividing the image information into one screen. Yes.

1 is a block diagram illustrating a configuration example of a first embodiment of an image processing device according to the present invention. It is a figure used in order to demonstrate the principal part of 1st Embodiment of the image processing apparatus by this invention. It is a figure used in order to demonstrate the effect of a 1st embodiment of the image processing device by this invention. It is a figure for demonstrating the example of a change process of the block division position in 1st Embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the example of a change process of the block division position in 1st Embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the example of a change process of the block division position in 1st Embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the example of a change process of the block division position in 1st Embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the example of a change process of the block division position in 1st Embodiment of the image processing apparatus by this invention. It is a figure used in order to demonstrate the block matching process in 2nd Embodiment of the image processing apparatus by this invention. It is a figure used in order to demonstrate the block matching process in 2nd Embodiment of the image processing apparatus by this invention. It is a figure used in order to demonstrate the block matching process in 2nd Embodiment of the image processing apparatus by this invention. It is a figure used in order to demonstrate the block matching process in 2nd Embodiment of the image processing apparatus by this invention. It is a figure used in order to demonstrate the block matching process in 2nd Embodiment of the image processing apparatus by this invention. It is a figure used in order to demonstrate the block matching process in 2nd Embodiment of the image processing apparatus by this invention. It is a figure used in order to demonstrate the block matching process in 2nd Embodiment of the image processing apparatus by this invention. It is a figure used in order to demonstrate the block matching process in 2nd Embodiment of the image processing apparatus by this invention. It is a figure used in order to demonstrate the block matching process in 2nd Embodiment of the image processing apparatus by this invention. It is a block diagram which shows the structural example of the imaging device with which 2nd Embodiment of the image processing apparatus by this invention is applied. It is a block diagram for demonstrating the block matching process in 2nd Embodiment of the image processing apparatus by this invention. It is a figure which shows the example of the captured image to which 2nd Embodiment of the image processing apparatus by this invention is applied. It is the figure which matched and showed on the image the local motion vector detected in 2nd Embodiment of the image processing apparatus by this invention. It is a figure used in order to demonstrate the principal part of 2nd Embodiment of the image processing apparatus by this invention. It is a figure used in order to demonstrate the principal part of 2nd Embodiment of the image processing apparatus by this invention. In the second embodiment of the image processing apparatus according to the present invention, local motion vectors that are considered to be highly reliable are shown in association with each other on the image. It is a block diagram for demonstrating the structural example of the motion vector calculation part which is the principal part of 2nd Embodiment of the image processing apparatus by this invention. It is a figure which shows the flowchart for demonstrating the main processing operations of 2nd Embodiment of the image processing apparatus by this invention. FIG. 26 is a diagram used for explaining the configuration example of FIG. 25. It is a flowchart for demonstrating the process example by the structural example of the principal part of 2nd Embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the formula used in order to demonstrate the example of a process which calculates | requires global motion from a local motion vector. It is a figure for demonstrating the formula used in order to demonstrate the example of a process which calculates | requires global motion from a local motion vector. It is a figure for demonstrating the formula used in order to demonstrate the example of a process which calculates | requires global motion from a local motion vector. It is a figure for demonstrating the formula used in order to demonstrate the example of a process which calculates | requires global motion from a local motion vector. It is a figure for demonstrating the formula used in order to demonstrate the example of a process which calculates | requires global motion from a local motion vector. It is a figure for demonstrating the problem of the conventional affine transformation. It is a figure used in order to explain other methods of calculating global motion. It is a figure for demonstrating the extended affine transformation used in 2nd Embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the formula used in order to demonstrate the process example which calculates a global motion from a local motion vector in 2nd Embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the formula used in order to demonstrate the process example which calculates a global motion from a local motion vector in 2nd Embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the formula used in order to demonstrate the process example which calculates a global motion from a local motion vector in 2nd Embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the formula used in order to demonstrate the process example which calculates a global motion from a local motion vector in 2nd Embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the formula used in order to demonstrate the process example which calculates a global motion from a local motion vector in 2nd Embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the formula used in order to demonstrate the process example which calculates a global motion from a local motion vector in 2nd Embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the formula used in order to demonstrate the process example which calculates a global motion from a local motion vector in 2nd Embodiment of the image processing apparatus by this invention. It is a figure which shows a part of flowchart for demonstrating the example of a process which calculates | requires global motion from a local motion vector in embodiment of 2nd Embodiment of the image processing apparatus by this invention. It is a figure which shows a part of flowchart for demonstrating the example of a process which calculates | requires global motion from a local motion vector in embodiment of 2nd Embodiment of the image processing apparatus by this invention. It is the figure which showed the motion vector of the block unit calculated | required from the calculated global motion in association with on the image in 2nd Embodiment of the image processing apparatus by this invention. In the second embodiment of the present invention, it is a diagram used to describe a process of evaluating the background matching degree of a target block using a global motion vector GMV. It is a figure for demonstrating the structural example of the background / moving subject determination part which is the principal part of 2nd Embodiment of this invention. It is a figure which shows a part of flowchart for demonstrating the processing operation in the background / moving subject determination part which is the principal part of 2nd Embodiment of this invention. It is a figure which shows a part of flowchart for demonstrating the processing operation in the background / moving subject determination part which is the principal part of 2nd Embodiment of this invention. It is a figure for demonstrating the other structural example of the background / moving subject determination part which is the principal part of 2nd Embodiment of this invention. It is a figure for demonstrating the further another structural example of the background / moving subject determination part which is the principal part of 2nd Embodiment of this invention. It is a figure for demonstrating the further another structural example of the background / moving subject determination part which is the principal part of 2nd Embodiment of this invention. It is a figure which shows a part of flowchart for demonstrating the processing operation in the further another structural example of the background / moving subject determination part which is the principal part of 2nd Embodiment of this invention. It is a figure which shows a part of flowchart for demonstrating the processing operation in the further another structural example of the background / moving subject determination part which is the principal part of 2nd Embodiment of this invention. It is a figure for demonstrating the further another structural example of the background / moving subject determination part which is the principal part of 2nd Embodiment of this invention. It is a block diagram which shows a part of example of a structure of an example of the addition rate calculation part which is the principal part of 2nd Embodiment of this invention. It is a block diagram which shows a part of example of a structure of an example of the addition rate calculation part which is the principal part of 2nd Embodiment of this invention. It is a figure used in order to demonstrate the example of a structure of an example of the addition rate calculation part which is the principal part of 2nd Embodiment of this invention. It is a figure used in order to demonstrate the example of a structure of an example of the addition rate calculation part which is the principal part of 2nd Embodiment of this invention. It is a figure used in order to demonstrate the example of a structure of an example of the addition rate calculation part which is the principal part of 2nd Embodiment of this invention. It is a figure used in order to demonstrate the example of a structure of an example of the addition rate calculation part which is the principal part of 2nd Embodiment of this invention. It is a figure used in order to demonstrate the example of a structure of an example of the addition rate calculation part which is the principal part of 2nd Embodiment of this invention. It is a figure used in order to demonstrate the example of a structure of an example of the addition rate calculation part which is the principal part of 2nd Embodiment of this invention. It is a figure used in order to demonstrate the example of a structure of an example of the addition rate calculation part which is the principal part of 2nd Embodiment of this invention. It is a figure used in order to explain the effect of a 2nd embodiment of this invention. It is a figure used in order to explain the effect of a 2nd embodiment of this invention. It is a figure used in order to explain the effect of a 2nd embodiment of this invention. It is a figure used in order to explain the effect of a 2nd embodiment of this invention. It is a figure used in order to explain the effect of a 2nd embodiment of this invention. It is a figure used in order to explain the effect of a 2nd embodiment of this invention. It is a figure for demonstrating the 1st process example in 3rd Embodiment of this invention. It is a figure for demonstrating the 1st process example in 3rd Embodiment of this invention. It is a figure for demonstrating the 2nd processing example in 3rd Embodiment of this invention. It is a figure for demonstrating the 2nd processing example in 3rd Embodiment of this invention. It is a figure which shows the flowchart for demonstrating the processing operation of the principal part in 4th Embodiment of this invention. It is a figure used in order to demonstrate the effect by a 4th embodiment of this invention. It is a figure used in order to demonstrate the effect by a 4th embodiment of this invention. It is a figure for demonstrating the other example of the process which calculates | requires a local motion vector in embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the other example of the process which calculates | requires a local motion vector in embodiment of the image processing apparatus by this invention. It is a figure for demonstrating the other example of the process which calculates | requires a local motion vector in embodiment of the image processing apparatus by this invention. It is a block diagram which shows the structural example of other embodiment of the image processing apparatus by this invention. It is a block diagram which shows the structural example of other embodiment of the image processing apparatus by this invention. It is a figure which shows the flowchart for demonstrating the processing operation of embodiment of FIG. It is a block diagram for demonstrating the structural example of the conventional image processing apparatus. It is a figure for demonstrating an example of a block division process. It is a figure for demonstrating the problem of the conventional image processing apparatus.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1000 ... Image processing part, 1001 ... Block division processing part, 1002 ... Motion vector detection part, 1003 ... Motion compensation image generation part, 1004 ... Image addition part, 1005 ... Division position shift control part, 1006 ... Timing signal generation part, 1010 Control unit, 1011 Block division processing unit, 1012 Image recognition processing unit, 15 Motion vector calculation unit, 16 Motion compensation image generation unit, 17 Addition unit

Claims (10)

Image processing means for performing predetermined image processing on the image in units of blocks obtained by dividing an image for one screen into a plurality of blocks;
The image for one screen is divided into a plurality of blocks, the block unit image is generated, and the generated block unit image information is passed to the image processing means, and the block on the one screen is generated. An image processing apparatus comprising: a block dividing unit that changes the divided position to a different position each time processing of an image for one screen is completed. The image processing apparatus according to claim 1.
The image processing means includes
Motion vector detecting means for detecting a motion vector for each block between a target image consisting of image information for one screen and a reference image consisting of one screen different from the target image;
Motion compensation means for generating a motion compensated image in which each block of the reference image is motion compensated using the motion vector for each block detected by the motion vector detection means;
An adding unit configured to superimpose the target image and the motion compensated image in units of blocks and to output an added image after the superimposition;
An image processing apparatus comprising: The image processing apparatus according to claim 2,
In the image processing means,
An image processing apparatus that superimposes two or more different reference images on the same target image. The image processing apparatus according to claim 2,
An image processing device that superimposes three or more different images using the added image as a target image. The image processing apparatus according to any one of claims 1 to 4,
The image processing apparatus, wherein the image information is captured image information transmitted from an image sensor at a predetermined frame rate. The image processing apparatus according to any one of claims 1 to 4,
The block dividing means includes
An image processing apparatus that changes a starting point of the block division at a ratio of a golden ratio within the range of the one block. The image processing apparatus according to any one of claims 1 to 6,
The block unit is an image processing apparatus in which an image of a rectangular area composed of 8 pixels in the horizontal direction and 8 lines in the vertical direction is used. In the image processing device according to any one of claims 1 to 7,
An image processing apparatus, wherein the block dividing means changes the block dividing position by changing the starting position of the divided block to one image unit or a plurality of pixel units. An image processing step of performing predetermined image processing on the image in units of blocks obtained by dividing an image for one screen into a plurality of images, and outputting post-processing image information for one screen;
Prior to the image processing step, the image for one screen is divided into a plurality of images, the block unit image is generated, and the generated block unit image information is passed to the image processing step. An image processing method comprising: a block dividing step of changing a block dividing position on one screen to a different position every time processing of an image for one screen is completed. The image processing method according to claim 9.
A motion vector detecting step for detecting a motion vector for each block between a target image consisting of image information for one screen and a reference image consisting of one screen different from the target image;
Using the motion vector for each block detected in the motion vector detection step, a motion compensation step for generating a motion compensated image in which each block of the reference image is motion-compensated;
An adding step for superimposing the target image and the motion compensated image in units of the blocks and outputting an added image after the superposition;
An image processing method comprising: