JP4286301B2 - Camera shake correction device, camera shake correction method, and recording medium recording camera shake correction program - Google Patents

Description

  The present invention relates to a camera shake correction device, a camera shake correction method, and a recording medium on which a camera shake correction program is recorded.

  2. Description of the Related Art Conventionally, a camera shake correction device that performs camera shake correction on an image captured by an imaging device such as a video camera is known. FIG. 1 is a block diagram of a camera shake correction apparatus. The camera shake correction apparatus includes a frame input unit 50, a geometric transformation coefficient calculation unit 51, a camera shake correction unit 52, and a frame output unit 53.

  For example, the geometric conversion coefficient calculation unit 51 calculates a geometric conversion coefficient from a plurality of frames using image processing such as matching. The camera shake correction unit 52 uses the geometric transformation coefficient calculated by the geometric transformation coefficient calculation unit 51 to geometrically transform the input frame once stored in the memory.

  The geometric transformation coefficient is represented by a 3 × 3 matrix as shown in the following equation (1), and corresponds to a world coordinate transformation coefficient between two frames.

  FIG. 2 shows a correction processing procedure in the case of performing camera shake correction by calculating a geometric transformation coefficient using image processing and geometrically transforming an input frame.

The starting frame (reference frame) is F _S , the current input frame is F ₂ , the input frame immediately before the current input frame F ₂ is F ₁ , and the geometric transformation coefficient between frames F ₁ and F ₂ (frame F ₂ The geometric transformation coefficient for converting the upper coordinates to the coordinates on the frame F ₁ ) is M ₁₂ , the geometric transformation coefficient between the frames F _S and F ₁ is M _S1 , and the cumulative geometric transformation coefficient between the frames F _S and F _{2 Assume} that M _S2 is a cumulative geometric conversion coefficient for converting coordinates on the frame F ₂ to coordinates on the frame F _S.

The following processing is performed on each frame from the start frame to the end frame. That is, (step 100) after clearing the memory for output frame, reads the frame on the work memory, the read frame (input frame) and F ₂ (Step 101).

Next, it is determined whether or not the input frame F ₂ is the start frame F _S (step 102). If it is determined that the input frame F ₂ is the start frame F _S , the cumulative geometric conversion coefficient M _S2 between the start frame F _S and the input frame F ₂ is determined to be no conversion as shown in Expression (2). A coefficient (initial value) to be expressed is set (step 103).

Then, the set input frame F ₂ as input frames F ₁ of the previous, set the M _S2 as the geometric conversion coefficient M _S1 between the input frames F ₁ start frame F _S and the previous (step 104). Further, the data of the input frame F ₂ is rendered as it is in the output frame memory (step 105). Thereafter, the process returns to step 100 to clear the output frame memory and read the next frame.

If it is determined in step 102 that the input frame F ₂ is not the start frame F _S , the geometric transformation coefficient M ₁₂ between the previous frame F ₁ and the current frame F ₂ is calculated (step 106). Next, as shown in Equation (3), an accumulated geometric transformation coefficient M _S2 between the start frame F _S and the current frame F ₂ is calculated from M _S1 and M ₁₂ (step 107).

Then, the set input frame F ₂ as input frames F ₁ of the previous, set the M _S2 as the geometric conversion coefficient M _S1 between the input frames F ₁ start frame F _S and the previous (step 108). Further, after the input frame F ₂ is geometrically transformed with the cumulative geometric transformation coefficient M _S2 , the frame after geometric transformation is rendered in the output frame memory (step 109). Thereafter, the process returns to step 100 to clear the output frame memory and read the next frame.

  A typical method for calculating the geometric transformation coefficient between two frames is shown below (see Japanese Patent Application Laid-Open No. 11-339021). First, feature points such as edge portions are extracted from one frame. Next, feature points are tracked (matched) between frames, and corresponding points on the other frame with respect to feature points on one frame are obtained. Since two equations are obtained from one set of corresponding points (x1, y1) and (x1 ′, y1 ′) as shown in equation (4), six or more equations are detected by detecting three or more sets of corresponding points. Is derived. From these equations, an unknown number of geometric transformation coefficients are calculated by a linear solution.

In the conventional geometric transformation coefficient calculation method, the cumulative geometric transformation coefficient between the start frame F _S and the input frame F ₂ is accumulated, and the geometric transformation coefficient between each adjacent frame between the start frame F _S and the input frame F ₂ is accumulated. As a result, the calculation error of the geometric transformation coefficient between adjacent frames accumulates as the number of input frames increases. In other words, when the geometric transformation coefficient is calculated incorrectly in the middle frame, there is a problem that the subsequent frames are corrected based on the frame corrected with the wrong transformation coefficient.

  Further, the conventional camera shake correction method has a problem that an unrendered area occurs in the output frame memory. That is, when the frames imaged in the presence of camera shake are frames as shown in FIGS. 3A to 3D, images obtained by rendering the data after geometric transformation of these frames in the output frame memory Is as shown in FIGS. 4 (a) to 4 (d). In FIG. 4B to FIG. 4D, the shaded area is an unrendered area.

Japanese Patent Laid-Open No. 11-339021 Japanese Patent No. 2892685

  According to the present invention, there is provided a camera shake correction apparatus, a camera shake correction method, and a recording medium on which a camera shake correction program is recorded that can prevent an unrendered area from occurring in an output frame memory area in which an image after camera shake correction is stored. The purpose is to provide.

  Further, the present invention avoids the occurrence of a non-rendered area in the output frame memory area in which the image after camera shake correction is stored, and can reduce the image quality deterioration while maintaining the angle of view. It is an object of the present invention to provide a correction apparatus, a camera shake correction method, and a recording medium on which a camera shake correction program is recorded.

  According to a first aspect of the present invention, in the camera shake correction apparatus that performs camera shake correction of moving image data, first means for calculating a geometric transformation coefficient between a frame serving as a reference for camera shake correction and an input frame for each input frame; And a second means for generating an output frame for each input frame, the second means clearing the output frame memory, and the geometric transformation corresponding to the input frame calculated by the first means. Means for geometric transformation based on a coefficient, means for rendering a frame after geometric transformation for an input frame in an output frame memory, and means for interpolating an unrendered area in the output frame memory with a plurality of frames near the input frame It is characterized by having.

  According to a second aspect of the present invention, in the camera shake correction method for performing camera shake correction of moving image data, a first step of calculating a geometric transformation coefficient between a frame serving as a reference for camera shake correction and the input frame for each input frame; And a second step of generating an output frame for each input frame, wherein the second step is a step of clearing the memory for the output frame, and a geometric transformation corresponding to the input frame calculated by the first step A step of performing geometric transformation based on a coefficient, a step of rendering a frame after geometric transformation for an input frame in an output frame memory, and a step of interpolating an unrendered region in the output frame memory by a plurality of frames in the vicinity of the input frame. It is characterized by.

  According to a third aspect of the present invention, there is provided a computer-readable recording medium on which a camera shake correction processing program for performing camera shake correction of moving image data is recorded, and a frame to be used as a reference for camera shake correction and an input frame for each input frame. A camera shake correction processing program for causing a computer to execute a first step of calculating a geometric transformation coefficient between the first step and a second step of generating an output frame for each input frame. Clearing the frame memory; geometrically transforming the input frame based on the corresponding geometric transformation coefficient calculated by the first step; rendering the geometrically transformed frame for the input frame in the output frame memory; And unrendered areas in the output frame memory Characterized in that it comprises a step of interpolating the plurality of frames in the vicinity of the input frame.

  According to the present invention, it is possible to avoid the occurrence of a non-rendered area in the output frame memory area in which the image after camera shake correction is stored.

  In addition, according to the present invention, it is possible to avoid the occurrence of a non-rendered area in the output frame memory area in which the image after camera shake correction is stored, and to reduce image quality degradation while maintaining the angle of view. become able to.

  Examples of the present invention will be described below.

[1] Description of Camera Shake Correction Device FIG. 5 shows a configuration of the camera shake correction device.

  The camera shake correction device is realized by a personal computer (PC) 10. A display 21, a mouse 22 and a keyboard 23 are connected to the PC 10. The PC 10 includes a CPU 11, a memory 12, a hard disk 13, and a removable disk drive (disk drive) 14 such as a CD-ROM.

  The hard disk 13 stores an image stabilization processing program in addition to an OS (operation system). The camera shake correction processing program is installed in the hard disk 13 using the CD-ROM 20 in which it is stored. Further, it is assumed that a moving image file captured by a video camera or the like is stored in the hard disk 13 in advance.

[2] Description of camera shake correction processing performed by the CPU when the camera shake correction program is started

  The camera shake correction process includes a geometric transformation coefficient calculation process and an output frame generation process. First, after a geometric transformation coefficient calculation process is performed, an output frame generation process is performed.

  In this embodiment, the user is allowed to set in advance whether or not to perform interpolation processing on an unrendered area in the output frame memory. When the setting for performing the interpolation process is performed on the unrendered area, a flag (hereinafter referred to as an interpolation process necessity determination flag) flag for storing the flag is set (flag = 1). In the case where the setting for performing the interpolation process on the unrendered area is performed, the interpolation process necessity determination flag flag for storing this is reset (flag = 0).

  In the camera shake correction process, the first frame (start frame) in the range in which camera shake correction is performed is used as a reference frame, and camera shake in other frames in the range in which camera shake correction is performed is corrected. Note that the range in which camera shake correction is performed can be arbitrarily set.

  FIG. 6 shows a geometric transformation coefficient calculation processing procedure.

F _S is the start frame (reference frame) of the range where camera shake correction is performed, F _{2 is} the current input frame within the range where camera shake correction is performed, F _{1 is} the input frame immediately before the current input frame F ₂ , and frame F The geometric transformation coefficient between ₁ and F ₂ (geometric transformation coefficient for transforming the coordinates on the frame F ₂ to the coordinates on the frame F ₁ ) is M ₁₂ , and the geometric transformation coefficient between the frames F _S and F ₁ is M _S1. The cumulative geometric transformation coefficient between frames F _S and F ₂ (the cumulative geometric transformation coefficient for converting the coordinates on the frame F ₂ to the coordinates on the frame F _S ) is M _S2 .

The following processing is performed on each frame from the start frame to the end frame in the range where camera shake correction is performed. That is, first, a frame is read into the memory, and the read frame (input frame) is set as F ₂ (step 1).

Next, it is determined whether or not the input frame F ₂ is the start frame F _S (step 2). When it is determined that the input frame F ₂ is the start frame F _S , there is no conversion as shown in the following equation (5) as the cumulative geometric conversion coefficient M _S2 between the start frame F _S and the input frame F _2. Is set (step 3).

Then, the set input frame F ₂ as input frames F ₁ of the previous, set the M _S2 as the geometric conversion coefficient M _S1 between the input frames F ₁ start frame F _S and the previous (step 4). Further, M _S2 is set as a direct geometric transformation coefficient M ′ _S2 between the start frame F _S and the input frame F ₂ (step 5). Thereafter, the process returns to step 1 to read the next frame.

If it is determined in step 2 that the input frame F ₂ is not the start frame F _S , the geometric transformation coefficient M ₁₂ between the previous frame F ₁ and the current frame F ₂ is calculated (step 6). Next, as shown in the following equation (6), a cumulative geometric transformation coefficient M _S2 between the start frame F _S and the current frame F ₂ is calculated from M _S1 and M ₁₂ (step 7).

Then, the set input frame F ₂ as input frames F ₁ of the previous, set the M _S2 as the geometric conversion coefficient M _S1 between the input frames F ₁ start frame F _S and the previous (step 8).

Next, a direct geometric transformation coefficient M ′ _S2 between the start frame F _S and the input frame F ₂ is calculated based on the cumulative geometric transformation coefficient M _S2 calculated in step 7 (step 9). Details of this processing will be described later. Thereafter, the process returns to step 1 to read the next frame. Such processing is repeated until the final frame.

  FIG. 7 shows the detailed procedure of the process in step 9 of FIG.

First, a plurality of feature points P ₂ are extracted from the input frame F ₂ (step 11). A feature point is a point where the luminance value of an adjacent pixel changes abruptly. And the process of steps 12-14 is performed for every feature point.

Read one feature point P ₂ (step 12). Then, the feature point P ₂ is subjected to coordinate transformation on the start frame F _S based on the cumulative geometric transformation coefficient M _S2 , and the point is set as P ′ ₂ (step 13). The feature point on the start frame F _S corresponding to the point P ₂ is traced using the point P ′ ₂ as an initial value of search (search start point), and the obtained point is set to P ″ ₂ (step 14).

Here, the tracking of feature points uses optical flow estimation by a gradient method. Since the optical flow estimation is premised on the movement of an object in a local region, if there is a large camera motion between the start frame F _S and the input frame F ₂ , the movement of the feature point cannot be accurately detected. Therefore, by setting the initial value of tracking on the start frame F _S and the P _'2 coordinate transformation cumulative geometric conversion coefficient M _S2 feature point P ₂ of the input frame F ₂ as described above, the start frame F _S even if there is a large camera motion between the input frame F _2, it is assumed to be a motion of minute, so the movement of the feature point can be accurately detected.

The processes in steps 12 to 14 are performed on all the feature points extracted in step 11, and the result of associating all the feature points between frames is obtained. As described above, two equations are derived from each corresponding point, and a geometric transformation coefficient of unknown 6 is calculated from all equations by a linear solution method (step 15). Let the calculated geometric transformation coefficient be M ′ _S2 .

  FIG. 8 shows another example of the processing procedure of step 9 in FIG.

First, a plurality of feature points P _S are extracted from the start frame F _S (step 21). A feature point is a point where the luminance value of an adjacent pixel changes abruptly. Next, the coefficient M _{S2 -inv} is obtained by inversely transforming the accumulated geometric transformation coefficient M _S2 (inverse matrix) (step 22). Thereafter, the processing of steps 23 to 25 is performed for each feature point.

Read one feature point P _S (step 23). Then, the feature point P _S is coordinate-transformed on the input frame F ₂ based on the coefficient M _{S2 -inv} , and the point is set as P ′ _s (step 24). Using the point P ′ _S as the initial value of the search (search start point), the feature point on the input frame F ₂ corresponding to the point P _S is tracked, and the obtained point is set as P ″ _S (step 25).

Here, the tracking of feature points uses optical flow estimation by a gradient method. Since the optical flow estimation is premised on the movement of an object in a local region, if there is a large camera motion between the start frame F _S and the input frame F ₂ , the movement of the feature point cannot be accurately detected. Accordingly, P ′ _S obtained by coordinate-transforming the feature point P _S of the start frame F _{S with} the coefficient M _S2-inv obtained by inversely transforming the cumulative geometric transformation coefficient M _S2 as described above is tracked on the input frame F _2. By using the initial value of, even if there is a large camera movement between the start frame F _S and the input frame F ₂ , it is assumed that the movement is minute, and the movement of the feature point can be accurately detected.

The processing in steps 23 to 25 is performed on all the feature points extracted in step 21 to obtain a result of associating all feature points between frames. As described above, two equations are derived from each corresponding point, and an unknown 6 geometric transformation coefficient is calculated from all the equations by a linear solution (step 26). Let the calculated geometric transformation coefficient be M ′ _S2 .

FIG. 9 shows an output frame generation processing procedure performed after calculating the geometric transformation coefficient M ′ _S2 for all frames.

The direct geometric transformation coefficient M ′ _S2 between the start frame F _S and each input frame F ₂ calculated by the geometric transformation coefficient calculation processing of FIG. 6 is _expressed as M ′ _Si (i = 1, 2,...). It will be expressed as

The following processing is performed on each frame from the start frame to the end frame. That is, first, the output frame memory is cleared (step 31). Thereafter, read target frame F _i in the memory (step 32). Next, the geometric transformation coefficient M ′ _Si corresponding to the target frame F _i is read (step 33). Next, a frame F ′ _i obtained by geometric transformation of the target frame F _i with the corresponding geometric transformation coefficient M ′ _Si is generated (step 34), and the frame F ′ _i after geometric transformation is rendered in the output frame memory (step 34). Step 35).

  Next, it is determined whether or not an interpolation processing necessity determination flag flag is set (flag = 1) (step 36). When the interpolation process necessity determination flag flag is not set (flag = 0), the interpolation process for each pixel in the unrendered area of the output frame memory (steps 37 to 41) is not performed. Therefore, the processes in steps 31 to 35 are performed for each frame from the start frame to the end frame.

  When the interpolation processing necessity determination flag flag is set, interpolation processing (steps 37 to 41) is performed for each pixel in the unrendered area of the output frame memory. An interpolation process for the pixel at the point P in the unrendered area will be described.

The 10 frames before and after the target frame F _i in time are extracted as the neighboring frames F _i-neighbors . If there are no 10 neighboring frames before and after, only limited neighboring frames are extracted. Then, the following processing is performed for each neighboring frame F _i-neighbors .

That is, the geometric transformation coefficient M ′ _Si-neighbors corresponding to the _neighboring frame F _i-neighbors is read into the memory (step 37). Next, a coefficient M ′ _{Si-neighbors-inv} obtained by _performing inverse transformation (inverse matrixing) of M ′ _Si-neighbors is obtained (step 38). Here, M ′ _{Si-neighbors-inv} corresponds to a geometric transformation coefficient for coordinate transformation of a point on the start frame onto the neighboring frame.

A point of the point P and geometric conversion in M _{'Si-neighbors-inv} and P _i-neighbors (step 39) to obtain the luminance value of P _i-neighbors points on the neighboring frames (step 40). Let the acquired luminance value be V _Pi-neighbors . Such processing (steps 37 to 40) is performed on each neighboring frame F _i-neighbors .

The luminance value of the point P is generated by mixing the luminance value V _Pi-neighbors corresponding to the point P obtained from each neighboring frame F _i-neighbors (step 41).

Note that when the point P _{i-neighbors obtained} by geometrically transforming the point P with M ′ _{Si-neighbors-inv} does not exist on the neighboring frame, the luminance values of the neighboring frame are not mixed. If there is no neighboring frame that can be used for interpolation, interpolation is performed with the pixel value of the start frame.

The mixing ratio of the luminance values V _Pi-neighbors corresponding to the point P obtained from each neighboring frame F _i-neighbors is determined as follows.

First, as shown in FIG. 10, a coefficient K of 1/2 ⁿ is given in advance to neighboring frames in order from the side closer to the target frame. n is the number of frames between the correction target frame and neighboring frames.

Next, the coefficients of neighboring frames that can be used for interpolation of the point P in the unrendered area are summed (K _total ). Then, the reciprocal of the sum (1 / K _total ) is calculated, and 1 / K _total is added to the coefficient K of each neighboring frame. This is the mixing ratio of each neighboring frame.

For example, if the neighboring frames that can be used for the interpolation of the point P in the unrendered region are one neighboring frame with K = 1/8 and one neighboring frame with K = 1/2, The total K _total of neighboring frame coefficients is 5/8. The reciprocal of this sum is 8/5. The mixing ratio of the neighboring frames with K = 1/8 is (1/8) × (8/5) = 1/5, and the mixing ratio of the neighboring frames with K = 1/2 is (1/2) × ( 8/5) = 4/5.

  In this way, by obtaining the mixture ratio with respect to neighboring frames that can be used for interpolation, the larger the mixture ratio is given to the pixel value of the neighboring frame that is temporally closer to the target frame, and the total of the mixture ratios is 1. Then, the interpolation processing for the point P as described above is performed for all the points in the unrendered area, thereby completing the output frame.

  According to the above embodiment, even if the calculation of the geometric transformation coefficient fails in the middle frame, the failure does not affect the correction processing of the subsequent frames, and the camera shake correction can be performed with high accuracy. In addition, since the output frame after camera shake correction does not include data other than shooting data, it is possible to secure a field of view that is greater than the input frame.

If it is not necessary to perform interpolation processing on the unrendered area, the output frame memory is cleared before step 1 in FIG. 6, and the frame F ₂ is geometrically transformed with M _s2 in step 5 in FIG. Then, the data after geometric transformation is rendered into an output frame, the geometric transformation coefficient M ′ _S2 is calculated in step 9 of FIG. 6 and the frame F ₂ is geometrically transformed with M ′ _s2 , and the data after geometric transformation is output into the output frame. You may make it render to.

  A modification of the output frame generation process will be described.

FIG. 11 shows another example of the output frame generation processing procedure performed after calculating the geometric transformation coefficient M ′ _S2 for all frames.

The direct geometric transformation coefficient M ′ _S2 between the start frame F _S and each input frame F ₂ calculated by the geometric transformation coefficient calculation processing of FIG. 6 is _expressed as M ′ _Si (i = 1, 2,...). It will be expressed as

  The following processing is performed on each frame from the start frame to the end frame. That is, first, it is determined whether or not an interpolation processing necessity determination flag flag is set (flag = 1) (step 41). When the interpolation process necessity determination flag flag is not set (flag = 0), the output frame memory is cleared (step 42), and then the process proceeds to step 43. When the interpolation process necessity determination flag flag is set (flag = 1), the process proceeds to step 43 without clearing the output frame memory.

In step 43, the target frame F _i is read into the memory. Next, the geometric transformation coefficient M ′ _Si corresponding to the target frame F _i is read (step 44). Next, a frame F ′ _i is generated by geometrically transforming the target frame F _i with the corresponding geometric transformation coefficient M ′ _Si (step 45), and the frame F ′ _i after the geometric transformation is rendered in the output frame memory (step 45). Step 46). The processes in steps 41 to 46 are repeated until the end frame.

  In this output frame generation process, when the interpolation processing necessity determination flag flag is set, the output frame memory is not cleared for each input frame. The camera shake correction result of each input frame is overwritten on the correction result. Since the start frame is not subjected to geometric transformation, after the start frame is rendered in the output frame memory, at least a region corresponding to the size of the start frame becomes a rendered region in the output frame memory.

  Therefore, when the camera shake correction result of each input frame is rendered in the output frame memory, the area other than the area where the camera shake correction result is rendered is interpolated with the camera shake correction result of the past frame. In this method, the output frame can secure the same field of view (angle of view) as the input frame. When there is no moving object in the periphery of the input frame, the joint between the image stabilization result of the previous frame and the image stabilization result of the current frame is relatively inconspicuous, and an image stabilization result with no sense of incongruity is obtained. It is done.

It is a block diagram which shows the structure of the conventional camera-shake correction apparatus. It is a flowchart which shows the camera-shake correction process sequence by a conventional method. It is a schematic diagram which shows the flame | frame imaged in the state with camera shake. FIG. 4 is a schematic diagram showing an image obtained by rendering data after geometric transformation of each frame in FIG. 3 in an output frame memory. It is a block diagram which shows the structure of a camera shake correction apparatus. It is a flowchart which shows the geometric transformation coefficient calculation processing procedure. It is a flowchart which shows the detailed process sequence of step 9 of FIG. It is a flowchart which shows the other example of the process sequence of step 9 of FIG. It is a flowchart which shows the production | generation process procedure of an output frame. It is a schematic diagram for demonstrating the calculation method of the mixture ratio of a near frame. It is a flowchart which shows the other example of the production | generation procedure of an output frame.

Explanation of symbols

10 Personal computer 11 CPU
12 Memory 13 Hard disk 14 Removable disk

Claims (3)

In a camera shake correction device that performs camera shake correction of moving image data,
For each input frame, a first means for calculating a geometric transformation coefficient between the input frame and a frame used as a reference for camera shake correction, and a second means for generating an output frame for each input frame,
The second means is means for clearing the output frame memory, means for geometrically transforming the input frame based on the corresponding geometric transformation coefficient calculated by the first means, and a frame after geometric transformation for the input frame as the output frame. A camera shake correction apparatus comprising: means for rendering in a memory; and means for interpolating an unrendered area in an output frame memory with a plurality of frames in the vicinity of an input frame. In the image stabilization method for correcting image stabilization of moving image data,
For each input frame, a first step of calculating a geometric transformation coefficient between a frame used as a reference for camera shake correction and the input frame, and a second step of generating an output frame for each input frame,
The second step is a step of clearing the output frame memory, a step of geometrically transforming the input frame based on the corresponding geometric transformation coefficient calculated in the first step, and a frame after geometric transformation with respect to the input frame as an output frame. A camera shake correction method comprising: rendering to a memory; and interpolating an unrendered region in an output frame memory with a plurality of frames in the vicinity of an input frame. A computer-readable recording medium recording a camera shake correction processing program for performing camera shake correction of moving image data,
For causing a computer to execute a first step of calculating a geometric transformation coefficient between a frame used as a reference for camera shake correction and an input frame for each input frame, and a second step of generating an output frame for each input frame We record camera shake correction processing program,
The second step is a step of clearing the output frame memory, a step of geometrically transforming the input frame based on the corresponding geometric transformation coefficient calculated in the first step, and a frame after geometric transformation with respect to the input frame as an output frame. A computer-readable recording medium having a camera shake correction processing program recorded thereon, comprising: a step of rendering in a memory; and a step of interpolating an unrendered area in an output frame memory with a plurality of frames near the input frame .

Images