JP2000134616A - Device and method for searching approximate image - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable high-speed processing eve when there is the positional relationship of rotation between reference source and destination images. SOLUTION: This device is provided with a reference source image memory part 1 for storing the image data of a reference source image 100, a reference destination image memory part 2 for storing the image data of a reference destination image 101, an image rotating part 6 capable of rotating the reference destination image 101 at plural rotating angles, a rotated image memory part 7 for storing the image data of a rotated reference destination image 114, an image selecting part 10 for selecting either a partial reference destination image 103 read out of the reference destination image memory part 2 or a partial reference destination image 117 from the rotated image memory part 7 and an image comparing part 11 for comparing a partial reference source image 102 from the reference source image memory part 1 with the partial reference destination image selected by the image selecting part 10 and at the image comparing part 11, a partial reference destination image 117 most approximate to the partial reference source image 102 is selected.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is suitable for an apparatus and method for encoding a still image or a moving image, and in particular, an approximate image search apparatus and method for searching for another partial area similar to a partial area of an image. About.

[0002]

2. Description of the Related Art As one of the techniques used for moving image coding, there is a frame prediction for predicting another frame image from a certain frame image using a motion vector representing the motion of an object. In this frame prediction, in order to detect a motion vector, an approximate image search for searching an image similar to a partial region of an image from another frame image is used.

In FIG. 25, in the case of image encoding using a motion vector, a reference source image 100 corresponds to a frame image encoded by moving image encoding. Reference image 10
Reference numeral 1 corresponds to another frame image encoded in the past in moving image encoding. The reference source image 100 is divided into a plurality of rectangular reference source partial images 102 called macroblocks. The image 103 is a rectangular image area (reference destination partial image) that is congruent with the reference source partial image 102 in the reference destination image 101.

In the motion vector search, each reference source partial image 102 is compared while moving the position of the reference destination partial image 103 in the reference destination image 101, and the most similar partial area is searched.

On the other hand, as a method of encoding a still image or a moving image, it is known that a partial area of an image is a reduced image of another partial area (self-similarity). Encoding). In fractal coding, a process of searching for another partial area similar to a certain partial area of an image in a reduced image of the image is required.

In FIG. 26, in the case of image encoding using fractal encoding, the reference source partial image 102 is called a range block (hereinafter, appropriately referred to as a range block 106), and is referred to as a reference original image 100 (hereinafter, appropriately, referred to as a range block 106). Range image 10
4 is assumed. ) Is divided into a plurality of parts by the reference source partial image 102. The rectangular partial image 107 in the reference source image 100 is called a domain block (hereinafter, appropriately referred to as a domain block 107), and is a rectangular partial image similar to the reference partial image 102. The size of the domain block 107 is arbitrary as long as it is larger than the range block 106.
It is assumed to be twice (in the horizontal and vertical directions, twice, that is, in the area, four times). In the fractal coding of an image, for each range block 106, a similar domain block 107 is searched from the same image. However, as an actual operation, instead of searching for a domain block 107 having a size twice as large as that of the range block 106, an image obtained by reducing the range image 104 to １ ／ is referred to as a reference image 101.
(Hereinafter, the domain image 105 is referred to as appropriate.), And the partial image 103 (the domain block 107 in the domain image 105) closest to the range block 106 is searched from the domain image 105.

In general, as an index indicating how close two images are, a sum E of square errors of respective pixels as shown in equation (1) is used.

[0008]

(Equation 1)

Note that p _i in equation (1) indicates the value of each pixel included in the reference source partial image 102, and q _i in equation (1) indicates each pixel included in the reference destination partial image 103. This is the pixel value. N in the expression is the number of pixels in each partial image.

In the approximate image search, the reference source image 10
0 is divided into several rectangular partial images 102, and a reference destination partial image 103 in which the value (sum E) of Expression (1) is the smallest for each partial image 102 is searched.

FIG. 27 shows how to represent image information in image encoding. The image size is m pixels wide,
Assuming that there are n pixels vertically, this image is treated as an image formed by collecting n lines having m pixels. When this image is handled by a computer or the like, the image information is held as a one-dimensional or two-dimensional data array on a memory.

FIG. 28 shows a case where an image is represented as a one-dimensional array. In this case, the image is handled as an array having m × n elements, and a pointer indicating the head thereof is added. Any pixel (x, y) in the image can be represented as the (x + y × m) th element of the array. The pixels adjacent to each other in the horizontal direction are also adjacent to each other on the memory, but the pixels adjacent to each other in the vertical direction are arranged at a distance from each other on the memory.

FIG. 29 shows a case where an image is represented as a two-dimensional array. In this case, each line of the image is handled as an array having m elements, and a pointer indicating the head of each array is an array of pointers. further,
A pointer indicating the head of the pointer array is added. Any pixel (x, y) in the image can be represented as the x-th element of the pixel data array pointed to by the y-th element of the pointer array. In this case as well, pixels adjacent in the horizontal direction are adjacent on the memory, but pixels adjacent in the vertical direction are arranged on the memory at distant positions.

[0014]

Normally, in image encoding, an approximate image search for a certain partial image is performed by searching only for an image which has a translational relationship with the partial image.
If an image having a relationship of not only translation but also rotation is added to the search candidates, the possibility of finding a better approximate image increases as the number of search candidates increases, and improvement in coding efficiency can be expected. .

However, as described with reference to FIGS. 27 to 29, in the image data stored in the image memory, pixels adjacent in the horizontal direction are adjacent in the memory.
Pixels that are adjacent in the vertical direction are arranged at distant positions in the memory, so that pixel data can be read at high speed when reading pixel data continuously from the image memory in the horizontal direction, but processing is performed when pixel data is read in the vertical direction. Speed drops.

In particular, a SIMD (Single Instruction Multipump) capable of processing a plurality of data with one instruction, such as the so-called MMX technology of an Intel microprocessor.
le data), the processing speed can be increased by performing calculations on multiple pixels at the same time.However, when loading the values of multiple pixels into a register, only the This is only for the pixels arranged in a line. For this reason, for example, when the reference source image and the reference destination image are in a rotational movement relationship, the image data cannot be efficiently read out to the register, and the calculation process is more complicated than in the case of the parallel movement relationship. It will be much slower.

FIG. 30 shows an example of calculating the difference between each pixel of the reference source partial image 102 and the reference destination partial image 103 using MMX technology of Intel Corporation as an example of the SIMD calculation function. Here, reference source partial image 1
02 and the size of the reference destination partial image 103 are 4 × 4 pixels. In FIG. 30, four pixels p _{0 to} p ₃ included in the top line of the reference source partial image 102 can be read into the register 108 by one instruction. Similarly, four pixels q included in the top line of the reference destination partial image 103
₀ to q ₃ can also be loaded into the register 109 with a single instruction. According to MMX technology, each of these registers 1
The difference calculation of each pixel held in 08 and 109, that is, p ₀ −q ₀ , p ₁ −q ₁ , p ₂ −q ₂ ,
The four calculations of p ₃ -q ₃ and the storage of each value of the operation result in the register 131 can be performed by one instruction.

FIG. 31 shows the reference destination partial image 103 of FIG.
Is rotated 90 degrees clockwise. At this time, the four pixels q _{0 to} q _{3 at} the left end of the reference destination partial image 103 in FIG. 31 cannot be read into the register 109 by one instruction. That is, in the case of FIG. 31, the four pixels q _{0 to} q _{3 at} the left end of the reference destination partial image 103 are pixel data arranged in the vertical direction on the image memory, and therefore, as in the reference destination partial image 103 of FIG. As compared with the case where pixel data arranged in a horizontal direction is read out on the image memory, it is necessary to execute extra memory access and operation instructions, and the execution speed is significantly reduced.

Therefore, the present invention has been made in view of such a situation, and even if there is a rotational positional relationship between the reference source image and the reference destination image, the execution speed of the operation instruction is not determined. Enables high-speed processing by suppressing deterioration,
An object of the present invention is to provide an approximate image search device and method.

[0020]

SUMMARY OF THE INVENTION An apparatus and method for searching for an approximate image according to the present invention searches for another partial image approximate to a partial area of an image, stores an image of a reference source, and stores an image of a reference destination. And stores an image obtained by rotating the reference destination image by a plurality of rotation angles, and retrieves the partial image extracted from the stored reference destination image and the rotated reference image stored from the stored reference destination image. The selected partial image, compares the partial image extracted from the stored reference source image with the selected partial image, calculates the difference between them, and selects the partial image closest to the reference source partial image By doing so, the problem described above is solved.

The apparatus and method for searching for an approximate image according to the present invention searches for another partial image that approximates a partial area of the image, stores an image of a reference source, stores an image of a reference destination, An image obtained by rotating the reference source image by a plurality of rotation angles is stored, and a partial image extracted from the stored reference source image and a partial image extracted from the rotated and stored reference source image are stored. By making a selection, comparing the partial image extracted from the stored reference destination image and the selected partial image, calculating the difference between them, and selecting the partial image that is closest to the reference source partial image, The above-mentioned problem is solved.

[0022]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described with reference to the drawings.

FIG. 1 shows a schematic configuration of an approximate image search device according to a first embodiment to which the approximate image search device and method of the present invention are applied.

The approximate image search apparatus according to the first embodiment shown in FIG. 1 includes a reference image memory unit 1 for storing image data of a reference image 100 and a reference image for storing image data of a reference image 101. Image memory unit 2 and reference destination image 101
, A rotated image memory unit 7 for storing image data of the rotated reference image 114, and a reference destination partial image 103 read from the reference image memory unit 2. Image data and rotated image memory unit 7
An image selection unit 10 for selecting any of the image data of the reference destination partial image 117 read out from the reference source image 102, the reference source partial image 102 from the reference source image memory unit 1, and the reference destination selected by the image selection unit 10. An image comparing unit 11 for comparing partial images is provided.

Hereinafter, the operation of the approximate image search apparatus of the first embodiment shown in FIG. 1 will be described. Here, as an example, a case will be described where the reference destination image 101 is rotated at four degrees of 0 degree (no rotation), 90 degrees, 180 degrees, and 270 degrees when searching for an approximate image.

In FIG. 1, the reference source image memory unit 1
The input image data of the reference source image 100 is stored.
Similarly, the reference destination image memory unit 2 stores the input image data of the reference destination image 101. The image rotation unit 6 generates a reference destination image 114 obtained by rotating the reference destination image 101,
The image data is developed in the rotation image memory unit 7. Further, the image selection unit 10 selectively selects one of the image data of the reference destination partial image 103 supplied from the reference destination image memory unit 2 and the image data of the reference destination partial image 117 supplied from the rotation image memory unit 7. And sends it to the image comparison unit 11.

Here, first, the image selecting section 10 retrieves each pixel data (image data with a rotation angle of 0 degree) of the reference destination partial image 103 from the reference destination image memory section 2 and sends it to the image comparison section 11. At this time, the image comparison unit 11 extracts each pixel data of the reference source partial image 102 from the reference source image memory unit 1, and square error with each pixel data of the reference destination partial image 103 sent from the image selection unit 10. Ask for. This operation is repeated for each of the plurality of reference destination partial images 103 included in the reference destination image 101, and the reference destination partial image 103 that gives the smallest square error (least square error) to the reference source partial image 102 is obtained.

Next, the image rotation unit 6 outputs the reference destination image 101.
Is rotated 90 degrees to create a reference destination image 114, and the image data is expanded in the rotated image memory unit 7. Further, the image selecting unit 10 extracts each pixel data of the reference destination partial image 117 from the rotated image memory unit 7 and sends it to the image comparing unit 11. At this time, the image comparison unit 11 calculates a square error between each pixel data of the reference source partial image 102 from the reference source image memory unit 1 and each pixel data of the reference destination partial image 117 sent from the image selection unit 10. Ask for. This operation is called reference image 1
14 is repeated for each of the plurality of reference destination partial images 117, and the smallest 2
Reference destination partial image 117 giving a squared error (least squared error)
Ask for. The above operation is similarly performed for the case where the rotation angles are 180 degrees and 270 degrees.

After that, the image comparing section 11 obtains a reference destination image giving the smallest square error from all the reference destination partial images 103 and 117 described above, and obtains its position information 123 and rotation angle information 124. Is output.

FIG. 2 is a flowchart showing the above processing.

In the flowchart shown in FIG.
In the approximate image search device of FIG. 1, first, as step S1, the image data of the reference source image 100 is input and stored in the reference source image memory unit 1, and at the same time, as step S2,
The image data of the reference destination image 101 is stored in the reference destination image memory unit 2
And store it.

Next, in step S3, the image comparison unit 11 stores the reference source image 10
2 are taken out. At this time, the image comparison unit 11 initializes the register Emin for holding the minimum value of the square error with a very large value in step S4. In the next step S5, the image selection unit 1
0, the reference destination partial image 10 from the reference destination image memory unit 2
3 and the image comparison unit 11 selects the reference source partial image 1
As described above, an approximate image search based on least squares error calculation is performed between 02 and the reference destination partial image 103.

In step S 6, the image rotation unit 6 sets the reference destination image 101 from the reference destination image memory unit 2.
Is rotated 90 degrees to generate a reference destination image 114, and the image data is stored in the rotated image memory unit 7. In the next step S7, the image selection unit 10 selects the reference destination partial image 117 from the rotated image memory unit 7, and the image comparison unit 11 determines between the reference source partial image 102 and the reference destination partial image 117 as described above. As described above, the approximate image search by the least square error calculation is performed.

Similarly, in step S 8, the image rotation unit 6 outputs the reference destination image 10 from the reference destination image memory unit 2.
A reference destination image 114 obtained by rotating 1 by 180 degrees is generated, and the image data is stored in the rotated image memory unit 7. In the next step S9, the image selection unit 10 selects the reference destination partial image 117 from the rotated image memory unit 7, and the image comparison unit 11 performs the above-described operation between the reference source partial image 102 and the reference destination partial image 117. As described above, the approximate image search by the least square error calculation is performed.

Further, in step S10, the image rotation section 6 causes the reference destination image 1 from the reference destination image memory section 2 to be read.
01 is rotated by 270 degrees to generate a reference destination image 114,
The image data is stored in the rotation image memory unit 7. In the next step S11, the image selection unit 10 selects the reference destination partial image 117 from the rotated image memory unit 7, and the image comparison unit 11 described above between the reference source partial image 102 and the reference destination partial image 117. As described above, the approximate image search by the least square error calculation is performed.

Next, in step S12, the image comparing section 11 obtains a reference destination image giving the smallest square error from all the reference destination partial images 103 and 117, and obtains the position information 123 and the rotation information. The angle information 124 is output.

After that, in step S13, the approximate image searching apparatus performs all the partial images 10 in the reference source image.
It is determined whether or not the processing for No. 2 has been completed. When it is determined that the processing has been completed, the processing of the approximate image search is terminated. When it is determined that the processing has not been completed, step S is performed.
The process returns to step 3.

The flowchart of FIG. 3 shows a process of searching for a reference destination partial image that is closest to the reference source partial image in steps S5, S7, S9, and S11 in the flowchart of FIG.

In the flowchart of FIG. 3, the image comparison unit 11 firstly refers to the reference destination partial image 103 or 117 from the reference destination image memory unit 2 or the rotated image memory unit 7 via the image selection unit 10 in step S21. Take out.

Next, in step S22, the image comparing section 11 makes the reference source partial image 102 and the reference destination partial image (10
3 or 117), the square error E using the above equation (1) is calculated.

Next, in step S23, the image comparison unit 11 sets the E in the register set in step S4 in FIG.
The magnitude comparison between min and E is performed. If Emin> E, the process proceeds to step S24. If Emin> E, the process proceeds to step S26.

In step S24, the image comparison section 11 sets Emin = E, and in step S25, records the position information 123 and the rotation angle information 124 of the reference destination partial image.

On the other hand, in step S26, the image comparing section 11 determines whether or not the processing has been completed for all the partial images of the reference destination image. When it is determined that the processing has not been completed, the processing returns to step S21, and the processing returns to step S21. When it is determined that the process has been completed, the process proceeds to the next stage of the flowchart of FIG.

In the above-described image comparison section 11, for example, the four pixels q _{0 at} the left end when the reference destination partial image 103 is rotated clockwise by 90 degrees as described with reference to FIG.
Although q ₃ cannot be loaded from the image memory unit at one time, in the approximate image search device according to the first embodiment of the present invention, the reference image 101 can be rotated in advance by the image rotation unit 6. Reference image 1 after rotation
Since the reference destination partial image 117 extracted from the image data 14 is used, the speed of access to the image memory unit is increased.

When the operation of the image comparison unit 11 in the first embodiment is realized by using the Intel MMX technology, it can be represented as shown in FIG.
In FIG. 4, components having functions similar to those in FIGS. 30 and 31 are denoted by the same reference numerals. In the example of FIG. 4, the size of the reference source partial image 102 and the reference destination partial images 103 and 117 is 4 × 4 pixels.

That is, in FIG. 4, four pixels p ₀ -p included in the top line of the reference source partial image 102
₃ can be read into the register 108 by one instruction. On the other hand, references section 4 pixel q ₀ to q ₃ of the leftmost If the image 103 is rotated 90 degrees clockwise, the referenced partial image 117 four pixels q ₀ horizontal row in on the top line
The ~q _3. Therefore, this reference destination partial image 117
4 pixel q ₀ to q ₃ of, can be read in the register 109 with a single instruction. Thereby, in the present embodiment, the difference calculation of each pixel held in each of the registers 108 and 109,
That is, p ₀ −q ₀ , p ₁ −q ₁ , p ₀
Four calculations _{2 -q 2, p 3 -q 3} , it becomes executable in one instruction for storing the value of the operation result to the register 131.

Next, FIG. 5 shows a schematic configuration of an approximate image search apparatus according to a second embodiment to which the approximate image search apparatus and method of the present invention are applied. In FIG. 5, components having the same functions as those in FIG. 1 are denoted by the same reference numerals.

The approximate image search apparatus according to the present embodiment includes a reference source image memory unit 1 for storing image data of a reference source image 100, a reference destination image memory unit 2 for storing image data of a reference destination image 101, and An image rotation unit 6 that can rotate the reference source image 100 at a plurality of rotation angles, and a rotated reference source image 2
And a reference source partial image 1 read from the reference source image memory unit 1.
02, the image selection unit 10 for selecting any of the image data of the reference source partial image 217 read from the rotated image memory unit 7, the reference destination partial image 103 from the reference destination image memory unit 2, and the image. An image comparing unit 11 for comparing the reference source partial images selected by the selecting unit 10 is provided.

Hereinafter, the operation of the approximate image searching apparatus according to the second embodiment shown in FIG. 5 will be described. In the second embodiment, an example will be described in which the reference source image is rotated at four degrees of 0 degree (no rotation), 90 degrees, 180 degrees, and 270 degrees when searching for an approximate image. I do.

In FIG. 5, the reference source image memory unit 1
The input image data of the reference source image 100 is stored.
Similarly, the reference destination image memory unit 2 stores the input image data of the reference destination image 101. The image rotation unit 6 generates a reference source image 214 obtained by rotating the reference source image 100,
The image data is developed in the rotation image memory unit 7. The image selection unit 10 selectively selects one of the image data of the reference source partial image 102 supplied from the reference source image memory unit 1 and the image data of the reference source partial image 217 supplied from the rotated image memory unit 7. And sends it to the image comparison unit 11.

Here, first, the image selection unit 10 extracts each pixel data (image data at a rotation angle of 0 °) of the reference source partial image 102 from the reference source image memory unit 1 and sends it to the image comparison unit 11. At this time, the image comparison unit 11 extracts each pixel data of the reference destination partial image 103 from the reference destination image memory unit 2, and square error with each pixel data of the reference source partial image 102 sent from the image selection unit 10. Ask for. This operation is repeated for each of the plurality of reference destination partial images 103 included in the reference destination image 101, and the reference destination partial image 103 that gives the smallest square error (least square error) to the reference source partial image 102 is obtained.

Next, the image rotation unit 6 outputs the reference source image 100
Is generated by rotating the reference source image 214 by 90 degrees, and the image data is expanded in the rotated image memory unit 7. Further, the image selecting unit 10 extracts each pixel data of the reference source partial image 217 from the rotated image memory unit 7 and sends the pixel data to the image comparing unit 11. At this time, the image comparison unit 11 squares each pixel data of each reference destination partial image 103 from the reference destination image memory unit 2 and each pixel data of the reference source partial image 217 sent from the image selection unit 10. Find the error. This operation is repeated for each of the plurality of reference destination partial images 103 included in the reference destination image 101, and the reference destination partial image 21 that gives the smallest square error (least square error) to the reference source partial image 217
Ask for 7. The above operation is performed at a rotation angle of 180 degrees and 270 degrees.
The same applies to the case of degrees.

Thereafter, the image comparing section 11 obtains the reference destination partial image 103 giving the smallest square error from all the reference destination partial images 103 described above, and obtains the position information 123 and the rotation angle information 124 thereof. Output.

FIG. 6 is a flowchart showing the above processing.

In the flowchart shown in FIG.
In the approximate image search device of FIG. 5, first, in step S31, the image data of the reference source image 100 is input to the reference source image memory unit 1 and stored, and at the same time, the image data of the reference destination image 101 is referred to in step S32. It is input to the destination image memory unit 2 and stored.

Next, in step S33, the image comparison unit 11 stores the reference source image 1
02 is taken out. At this time, in step S34, the image comparison unit 11 initializes a register Emin for holding the minimum value of the square error with a very large value. In the next step S35, the image selection unit 10 selects the reference source partial image 102 from the reference source image memory unit 1, and the image comparison unit 11 compares the reference source partial image 102 and the reference destination partial image 103 with each other. As described above, the approximate image search by the least square error calculation is performed.

In step S 36, the image rotation unit 6 outputs the reference source image 10 from the reference source image memory unit 1.
The reference source image 214 obtained by rotating 0 by 90 degrees is generated, and the image data is stored in the rotated image memory unit 7. In the next step S37, the image selection unit 10 selects the reference source partial image 217 from the rotated image memory unit 7, and the image comparison unit 11 performs the above-mentioned operation between the reference destination partial image 103 and the reference source partial image 217. As described above, the approximate image search by the least square error calculation is performed.

Similarly, in step S 38, the image rotation unit 6 reads the reference source image 1 from the reference source image memory unit 1.
Generate a reference source image 214 obtained by rotating 00 by 180 degrees,
The image data is stored in the rotation image memory unit 7. In the next step S39, the image selection unit 10 selects the reference source partial image 217 from the rotated image memory unit 7, and the image comparison unit 11 performs the above-described operation between the reference destination partial image 103 and the reference source partial image 217. As described above, the approximate image search by the least square error calculation is performed.

Further, in step S40, the image rotation unit 6 causes the reference source image 1 from the reference source image memory unit 2 to be read.
Generate a reference source image 214 obtained by rotating 00 by 270 degrees,
The image data is stored in the rotation image memory unit 7. In the next step S41, the image selection unit 10 selects the reference source partial image 217 from the rotated image memory unit 7, and the image comparison unit 11 performs the above-described processing between the reference destination partial image 103 and the reference source partial image 217. As described above, the approximate image search by the least square error calculation is performed.

Next, in step S42, the image comparing section 11 obtains the reference partial image 103 giving the smallest square error from all the reference partial images 103, and obtains the position information 123 and the rotation angle. The information 124 is output.

Thereafter, in step S43, the approximate image searching apparatus performs all the partial images 10 in the reference source image.
It is determined whether the processing for 2 and 217 has been completed,
When it is determined that the processing is completed, the processing of the approximate image search is terminated, and when it is determined that the processing is not completed, the processing returns to step S33.

In the second embodiment, the same effect as in the first embodiment is obtained by rotating the reference source image instead of rotating the reference destination image as in the first embodiment. Can be obtained.

Next, FIG. 7 shows a schematic configuration of an approximate image search device according to a third embodiment to which the approximate image search device and method of the present invention are applied. Note that, in FIG. 7, the same reference numerals are given to components having the same functions as those in the respective drawings.

The approximate image search device according to the third embodiment shown in FIG. 7 includes a reference source image memory unit 1 for storing image data of a reference source image 100, and a reference destination for storing image data of a reference destination image 101. Image memory unit 2 and reference destination image 101
An image rotating unit 6 that can rotate the image at a plurality of rotation angles, a plurality of rotated image memory units 7a, 7b, and 7c that respectively store image data of the rotated plurality of reference destination images 114a, 114b, and 114c; Image data of the reference destination partial image 103 read from the image memory unit 2 and each reference destination partial image 117 from the rotated image memory units 7a, 7b, 7c
The image selection unit 10 for selecting any of the image data a, 117b, and 117c, the reference source partial image 102 from the reference source image memory unit 1, and the reference destination partial image selected by the image selection unit 10 are compared. An image comparison unit 11 is provided.

Hereinafter, the operation of the approximate image searching apparatus of the third embodiment shown in FIG. 7 will be described. In the third embodiment, when an approximate image is searched, the reference destination image 101 is shifted by 0 degree (no rotation), 90 degrees, 180 degrees, and 2 degrees.
The description will be made by taking as an example a case where the rotation is performed in four ways of 70 degrees.

In FIG. 7, the reference source image memory section 1
The input image data of the reference source image 100 is stored.
Similarly, the reference destination image memory unit 2 stores the input image data of the reference destination image 101. The image rotation unit 6 generates a plurality of reference destination images 114a, 114b, 114c by rotating the reference destination image 101 by 90 degrees, 180 degrees, and 270 degrees, and stores the image data in the corresponding rotation image memory units 7a, 7b and 7c. The image selection unit 10
Are the image data of the reference destination partial image 103 supplied from the reference destination image memory unit 2 and the rotation image memory units 7a and 7
b, 7c each reference destination partial image 117a, 1
One of the image data 17b and 117c is selectively extracted and sent to the image comparison unit 11.

That is, the image selecting unit 10 extracts each pixel data (image data with a rotation angle of 0 degree) of the reference destination partial image 103 from the reference destination image memory unit 2 and sends it to the image comparison unit 11. Similarly, the image selecting unit 10 outputs the reference destination partial images 117a, 117a, 117a, and 117b from the rotated image memory units 7a, 7b, and 7c.
Each pixel data of 117 b and 117 c is extracted and sent to the image comparison unit 11.

At this time, the image comparison unit 11 compares each pixel data of the reference source partial image 102 from the reference source image memory unit 1 with the reference destination partial image 10 sent from the image selection unit 10.
The square errors with the respective pixel data of 3, 117a, 117b, and 117c are obtained. This operation is referred to as reference image 10
The reference partial image which gives the smallest square error (least square error) to the reference source partial image 102 is obtained by repeating for each of the plurality of reference destination partial images included in 1.

After that, the image comparing section 11 executes all the reference destination partial images 103 and 117a, 117b, and 11 described above.
7c, a reference destination partial image giving the smallest square error is obtained, and its position information 123 and rotation angle information 1 are obtained.
24 is output.

FIG. 8 is a flowchart showing the above processing.

In the flowchart shown in FIG.
In the approximate image search device of FIG. 7, first, in step S51, the image data of the reference source image 100 is input and stored in the reference source image memory unit 1, and at the same time, in step S52, the image data of the reference destination image 101 is referred to. It is input to the destination image memory unit 2 and stored.

Next, in step S 53, the image rotation section 6 reads the reference destination image 10 from the reference destination image memory section 2.
1 is rotated 90 degrees to generate a reference destination image 114a, and the image data is stored in the rotated image memory unit 7a. Similarly, in step S54, the image rotating unit 6 rotates the reference image 101 from the reference image memory unit 2 by 180 degrees to generate the reference image 114b, and stores the image data in the rotated image memory unit 7b. Then, in step S55, the reference destination image 101 from the reference destination image memory unit 2 is rotated by 270 degrees to generate the reference destination image 114c, and the image data is stored in the rotated image memory unit 7c.

Next, in step S56, the image comparison unit 11 stores the reference source image 1
02 is taken out. At this time, the image comparison unit 11 initializes the register Emin for holding the minimum value of the square error with a very large value in step S57.

In the next step S58, the image comparison section 11 makes the reference partial image 102 and the reference partial image 1
03, an approximate image search is performed by the least square error calculation as described above. Similarly, the image comparison unit 11 performs an approximate image search by a least square error operation between the reference source partial image 102 and the reference destination partial image 117a rotated by 90 degrees in step S59, and in step S60, performs a reference image search. Approximate image search by least squares error calculation is performed between 102 and the reference destination partial image 117b rotated by 180 degrees.
Approximate image search by least squares error calculation is performed between the reference image 117c rotated by 270 degrees.

Next, in step S62, the image comparison section 11 makes all the reference destination partial images 103 and 117a,
A reference destination partial image that gives the smallest square error is obtained from 117b and 117c, and its position information 123 and rotation angle information 124 are output.

Thereafter, in step S63, the approximate image searching apparatus performs all the partial images 10 in the reference source image.
It is determined whether or not the processing for No. 2 has been completed. When it is determined that the processing has been completed, the processing of the approximate image search is terminated. When it is determined that the processing has not been completed, step S is performed.
It returns to the process of 53.

In the third embodiment, the same effects as in the first embodiment can be obtained.

Next, FIG. 9 shows a schematic configuration of an approximate image search apparatus according to a fourth embodiment to which the approximate image search apparatus and method of the present invention are applied. In FIG. 9, components having the same functions as those in the respective drawings are given the same reference numerals.

The approximate image search apparatus according to the fourth embodiment shown in FIG. 9 includes a reference source image memory unit 1 for storing image data of a reference source image 100, and a reference destination for storing image data of a reference destination image 101. Image memory unit 2 and reference source image 100
An image rotation unit 6 that can rotate the image at a plurality of rotation angles, a plurality of rotated image memory units 7a, 7b, and 7c that store respective image data of a plurality of rotated reference source images 214a, 214b, and 214c, and a reference destination. The image data of the reference destination partial image 103 read from the image memory unit 2 and the respective reference source partial images 217 from the rotated image memory units 7a, 7b, 7c
a, 217b, and 217c are compared with the image selection unit 10 for selecting any of the image data, the reference destination partial image 103 from the reference destination image memory unit 2, and the reference source partial image selected by the image selection unit 10. An image comparison unit 11 is provided.

Hereinafter, the operation of the approximate image searching apparatus of the fourth embodiment shown in FIG. 9 will be described. In the fourth embodiment, when searching for an approximate image, the reference source image 100 is rotated by 0 degree (no rotation), 90 degrees, 180 degrees, and 2 degrees.
The description will be made by taking as an example a case where the rotation is performed in four ways of 70 degrees.

In FIG. 9, the reference source image memory unit 1
The input image data of the reference source image 100 is stored.
Similarly, the reference destination image memory unit 2 stores the input image data of the reference destination image 101. The image rotation unit 6 generates a plurality of reference source images 214a, 214b, 214c by rotating the reference source image 100 by 90 degrees, 180 degrees, and 270 degrees, and stores the image data in the corresponding rotated image memory units 7a, 7b and 7c. The image selection unit 10
Are the image data of the reference source partial image 102 supplied from the reference source image memory unit 2 and the rotation image memory units 7a and 7
b, 7c, the respective reference source partial images 217a, 217
One of the image data 17b and 217c is selectively extracted and sent to the image comparison unit 11.

That is, the image selecting unit 10 extracts each pixel data (image data with a rotation angle of 0 degrees) of the reference source partial image 102 from the reference source image memory unit 1 and sends it to the image comparison unit 11. Similarly, the image selecting unit 10 outputs the reference source partial images 217a, 217a, 217a, and 217a from the rotated image memory units 7a, 7b, and 7c.
Each pixel data of 217b and 217c is extracted and sent to the image comparison unit 11.

At this time, the image comparison unit 11 compares each pixel data of the reference destination partial image 103 from the reference destination image memory unit 2 with the reference source partial image 10 sent from the image selection unit 10.
A square error with each pixel data of 2,217a, 217b, and 217c is obtained. This operation is referred to as reference image 10
1 is repeated for each of the plurality of reference destination partial images 103 included in the reference source partial images 102, 217a, 217b,
Smallest square error for 217c (least square error)
Is obtained.

Thereafter, the image comparing section 11 obtains the reference destination partial image 103 giving the smallest square error from all the reference destination partial images 103 described above, and obtains the position information 123 and the rotation angle information 124 thereof. Output.

FIG. 10 is a flowchart showing the above processing.

In the flowchart shown in FIG. 10, the approximate image searching apparatus shown in FIG.
Then, the image data of the reference source image 100 is input and stored in the reference source image memory unit 1, and at the same time, as step S72, the image data of the reference destination image 101 is input and stored in the reference destination image memory unit 2.

Next, in step S 73, the image rotation section 6 reads the reference source image 10 from the reference source image memory section 1.
A reference source image 214a obtained by rotating 0 by 90 degrees is generated, and the image data is stored in the rotated image memory unit 7a. Similarly, in step S74, the image rotation unit 6 generates a reference source image 214b obtained by rotating the reference source image 100 from the reference source image memory unit 1 by 180 degrees, and stores the image data in the rotated image memory unit 7b. Then, in step S75, a reference source image 214c obtained by rotating the reference source image 100 from the reference source image memory unit 1 by 270 degrees is generated, and the image data is stored in the rotated image memory unit 7c.

Next, in step S76, the image comparing section 11 sends the reference source partial images 102, 21 from the reference source image memory section 1 and the rotated image memory sections 7a, 7b, 7c.
Each pixel data of 7a, 217b, and 217c is extracted.
At this time, the image comparison unit 11 determines in step S77 that the register Emi holds the minimum value of the square error.
Initialize n with a very large value.

In the next step S78, the image comparing section 11 makes the reference partial image 102 and the reference partial image 1
03, an approximate image search is performed by the least square error calculation as described above. Similarly, in step S79, the image comparison unit 11 performs an approximate image search by a least square error calculation between the reference destination partial image 103 and the reference source partial image 217a rotated by 90 degrees. An approximate image search is performed between the image 103 and the reference source partial image 217b rotated by 180 degrees by the least square error calculation. In step S81, the reference destination partial image 10
An approximate image search is performed between the 3 and the reference source partial image 217c rotated by 270 degrees by a least square error calculation.

Next, in step S82, the image comparing section 11 obtains the reference partial image 103 giving the smallest square error from all the reference partial images 103, and obtains the position information 123 and the rotation angle. The information 124 is output.

After that, in step S83, the approximate image searching apparatus performs all the partial images 10 of the reference source image.
2, 217a, 217b, and 217c are determined to be completed. If it is determined that the process is completed, the approximate image search process is terminated. If it is determined that the process is not completed, the process of step S73 is performed. Return to

In the fourth embodiment, the same effect as in the first embodiment is obtained by rotating the reference source image instead of rotating the reference destination image as in the third embodiment. Can be obtained.

In the first to fourth embodiments described above, the description has been made without specifying any of the image coding using the motion vector and the image coding using the fractal coding. In the case of the used image coding, FIG.
As shown in the above, the reference source image 100 corresponds to a frame image to be encoded, the reference destination image 101 is a previously encoded frame image, the reference source partial image 102 is a macro block, The reference destination partial image 103 corresponds to a macroblock in a previously encoded frame. The image coding using this motion vector can be applied to any of the first to fourth embodiments described above, and the configuration and operation of each are the same as those in FIGS.

On the other hand, in the first to fourth embodiments, in the case of image coding using fractal coding,
As shown in FIG. 26, the reference source image 100 is a range image 104, and the reference destination image 101 is the same image as the range image 104. The reference source partial image 102 is a range block 106. Also in the case of image coding using this fractal coding,
Although the present invention can be applied to any of the fourth embodiments, when the fractal coding is used, a 1/2 reduced image of the range image 104 is generated as the domain image 105. A reduced image generation unit for generating the image 105 is provided. Further, the reference destination partial image 103 becomes a domain block 107 in the domain image 105.

FIG. 11 shows, as a fifth embodiment of the present invention, an approximate image when the same function as that of the first embodiment is applied to image coding using fractal coding. 1 shows a schematic configuration of a search device. Note that, in FIG. 11, the same reference numerals are given to components having the same functions as those in the respective drawings.

The approximate image search device of the fifth embodiment shown in FIG. 11 includes a range image memory unit 3 for storing image data of a range image 104, and a reduced image generation unit for generating a domain image 105 from the range image 104. 13, a domain image memory unit 4 for storing image data of the domain image 105, an image rotation unit 6 capable of rotating the domain image 105 at a plurality of rotation angles, and a rotated image for storing image data of the rotated domain image 314. Memory unit 7 and domain block 10 read from domain image memory unit 4
7 and the image data of the domain block 317 from the rotated image memory unit 7, the image selecting unit 10, the range block 106 from the range image memory unit 3, and the image selecting unit 10. An image comparing unit 51 that compares the selected domain block with the selected domain block is provided.

Hereinafter, the operation of the approximate image searching apparatus of the fifth embodiment shown in FIG. 11 will be described. This fifth
In the embodiment, when the approximate image is searched, the domain image 105 is set to 0 degree (no rotation), 90 degrees, 180 degrees, and 27 degrees.
A description will be given of a case where the rotation is performed in four ways of 0 degrees.

In FIG. 11, the range image memory unit 3
Stores the image data of the range image 104 that is the input reference source image. The reduced image generation unit 13 generates the domain image 105 by reducing the range image 104 by half. The domain image memory unit 4 includes the reduced image generation unit 1
The domain image 105 generated in step 3 is stored. The image rotation unit 6 generates a domain image 314 obtained by rotating the domain image 105, and expands the image data in the rotated image memory unit 7. Further, the image selecting unit 10 selectively extracts either the image data of the domain block 107 supplied from the domain image memory unit 4 or the image data of the domain block 317 supplied from the rotated image memory unit 7 and compares the image data. It serves to send to the unit 51.

Here, the image selecting unit 10 extracts each pixel data (image data with a rotation angle of 0 degree) of the domain block 107 from the domain image memory unit 4 and sends it to the image comparing unit 51. At this time, the image comparison unit 51 extracts each pixel data of the range block 106 from the range image memory unit 3 and calculates a square error with each pixel data of the domain block 107 sent from the image selection unit 10 as described later. Ask. This operation is repeated for each of the plurality of range blocks 106 included in the range image 104, and a domain block 107 that gives the smallest square error (least square error) to the range block 106 as described later is obtained.

Next, the image rotation unit 6 converts the domain image 10
A domain image 314 obtained by rotating 5 by 90 degrees is created, and the image data is expanded in the rotated image memory unit 7. Further, the image selecting unit 10 extracts each pixel data of the domain block 317 from the rotated image memory unit 7 and
Send to 1. At this time, the image comparing unit 51 compares the pixel data of the range block 106 from the range image memory unit 3 with the domain block 31 sent from the image selecting unit 10.
The square error with each pixel data of No. 7 is obtained as described later. This operation is repeated for each of the plurality of domain blocks 317 included in the domain image 314, and a domain block 317 that gives the smallest square error (least square error) to the range block 106 is obtained as described later. The above operation is performed at a rotation angle of 180 degrees and 270 degrees.
The same applies to the case of degrees.

After that, the image comparing section 51 finds a domain block giving the smallest square error from all the domain blocks 107 and 317 described above, and determines the position and the rotation angle of the domain block by affine mapping. 120
Output as Also, the contrast 121 and the luminance offset 12 between the range block and the domain block
The information of No. 2 is similarly output.

FIG. 12 is a flowchart showing the above processing.

In the flowchart shown in FIG. 12, the approximate image searching apparatus shown in FIG.
In step S92, the image data of the range image 104 is input to the range image memory unit 3 and stored.
In the reduced image generating unit 13, the range image 1
04 is reduced to 1/2 to generate a domain image 105,
It is input to the domain image memory unit 4 and stored.

Next, in step S93, the image comparing section 51 stores the range block 1 from the range image memory section 3.
06 of each pixel data is extracted. At this time, the image comparison unit 51 initializes a register Emin for holding the minimum value of the square error with a very large value in step S94. In the next step S95, the image selecting unit 10 selects the domain block 107 from the domain image memory unit 4, and the image comparing unit 51 determines the least square error between the range block 106 and the domain block 107 as described later. An approximate image search is performed by calculation.

In step S96, the image rotation unit 6 generates a domain image 314 obtained by rotating the domain image 105 from the domain image memory unit 4 by 90 degrees, and stores the image data in the rotated image memory unit 7. . In the next step S97, the image selecting unit 10 selects the domain block 317 from the rotated image memory unit 7, and the image comparing unit 51 calculates the later-described least square error between the range block 106 and the domain block 317. Approximate image search is performed.

Similarly, in step S98, the image rotating unit 6 generates a domain image 314 obtained by rotating the domain image 105 from the domain image memory unit 4 by 180 degrees, and stores the image data in the rotated image memory unit 7. Let it. In the next step S99, the image selecting unit 10 selects the domain block 317 from the rotated image memory unit 7, and the image comparing unit 51 performs a later-described least square error calculation between the range block 106 and the domain block 317. Perform an approximate image search.

Further, in step S100, the image rotation unit 6 generates a domain image 314 obtained by rotating the domain image 105 from the domain image memory unit 4 by 270 degrees, and stores the image data in the rotated image memory unit 7. . In the next step S101, the image selection unit 10
Now, the domain block 317 from the rotated image memory unit 7
Is selected, and the image comparison unit 51 performs an approximate image search between the range block 106 and the domain block 317 by a least square error calculation described later.

Next, in step S102, the image comparison section 61 checks all the domain blocks 107 and 317.
From among the above, a domain block that gives the smallest square error is obtained, and the affine mapping 120 and the contrast 121
And information of the luminance offset 122 are output.

Thereafter, in step S103, the approximate image search device determines whether or not the processing has been completed for all the range blocks 106 in the range image 104. When the process is terminated and it is determined that the process is not completed, the process returns to step S93.

FIG. 13 is a flowchart showing steps S95, S97, S99 and S1 in the flowchart of FIG.
13 shows a process of searching for a domain block closest to a range block in FIG.

In the flowchart of FIG. 13, the image comparison unit 51 firstly performs the domain block 107 or 31 from the domain image memory unit 4 or the rotation image memory unit 7 via the image selection unit 10 in step S111.
Take out 7.

Next, in step S112, the image comparing section 51 calculates a contralast C between the range block 106 and the domain block (107 or 317) as described later. Calculate the luminance offset B.

Next, in step S114, the image comparing section 51 calculates the sum E of the square errors between the range block and the domain block by using equation (2) described later, and furthermore, in step S115. The magnitude of Emin of the register set in step S94 of FIG. 12 is compared with the magnitude of E, and if Emin> E, the process proceeds to step S94.
The process proceeds to the process of step 116, and if Emin> E is not satisfied, the process proceeds to the process of step S118.

In step S116, the image comparing section 51
Is set to Emin = E, and the position, rotation angle (affine mapping), contrast C and luminance offset B of the domain block are recorded in the next step S117.

On the other hand, in step S118, the image comparing section 51 determines whether or not processing has been completed for all domain blocks of the domain image. When it is determined that processing has not been completed, the processing returns to step S111, and processing is completed. If it is determined that it has been performed, the process proceeds to the next stage of the flowchart of FIG.

Also in the fifth embodiment, the same effects as in the above embodiments can be obtained.

Here, in the case of image coding using fractal coding, the image comparison section 51 calculates the sum of square errors E as in the following equation (2) instead of the above equation (1). calculate.

[0118]

(Equation 2)

[0119] Incidentally, r _i in the formula (2) represents a value of each pixel in the range block, d _i to the value of each pixel in a domain block, N is the number of pixels in the block, C is a range block B indicates a contrast between the domain block and B, and B indicates a luminance offset between the range block and the domain block.

The contrast C and the luminance offset B that minimize the sum E of the square errors are obtained by partially differentiating the equation (2) with C and B as shown in equations (3) and (4). It can be obtained from Expressions (3) and (4) as in Expressions (5) and (6).

[0121]

(Equation 3)

Further, the sum E of the square errors can be obtained from Expressions (2) and (6) as in Expression (7).

[0123]

(Equation 4)

However, the above equations (3), (4) and (7)
More, the contrast C, for calculating a luminance offset B and the square error sum E, the sum of the total sum of the pixel values within the range block? R _i and the sum of squares? R _i ^2, the pixel values in the domain block [Sigma] d _i and [Sigma] d _i ² needs to be calculated. Since this calculation includes many multiplications and divisions, the calculation amount is enormous if it is calculated for all range blocks and domain blocks. To reduce this, the image comparison unit 51
, The range image 104 and the domain image 10
5 than r _i, are stored to create a table of cumulative values of r _i ^2, d _i and d _i ^2.

The table of the accumulated values will be described with reference to FIG. The table 110 in FIG. 14 is a table of the accumulated pixel values of the given image.

In FIG. 14, assuming that the size of the image is m horizontal pixels × n vertical lines, the size of the cumulative value table 110 is horizontal (m + 1) × vertical (n + 1). Arbitrary element 112 (Cum (x, y)) in this table 110 is in the upper left area 1
It holds the sum of the pixel values included in the area corresponding to 11 (coordinates (0,0)-(x-1, y-1)). In particular, the lower right element 11
3 (Cum (m, n)) holds the sum of the pixel values of the entire screen (coordinates (0,0)-(m-1, n-1)). Using this table, the sum of pixel values in an arbitrary rectangular area (coordinates (x0, y0)-(x1-1, y1-1)) in the image can be obtained by the following equation (8).

[0127]

(Equation 5)

The sum of squares can be calculated by preparing a similar table. Since the cumulative value of the pixel values in the block and the cumulative square value are constant irrespective of the rotation state of the block, this table can be applied not only to the non-rotated domain image but also to the rotated domain image. That is, if the cumulative value table is created only once at the initialization stage, the sum of the pixel values and the sum of the squares can be obtained only by addition and subtraction, and the total amount of calculation can be greatly reduced.

FIG. 15 shows the internal structure of the image comparison unit 51 in FIG. 11 which realizes the above-described operation.

In FIG. 15, range block 10
6 is read by referencing the memory access unit 14, it is sent to the cumulative value table memory 16 and the? R _i ² cumulative value memory 17 for calculation for? R _i calculation. Similarly, domain blocks 107 and 317 are sent to the cumulative value table memory 18 and [Sigma] d _i ² cumulative value table memory 19 for calculation for [Sigma] d _i calculated by reference destination memory access unit 15. .Sigma.r _i d _i calculator 20, the reference source memory access unit 14, from the referenced memory access unit 15 reads the data, calculates the cumulative values of r _i × d _i.

The contrast C calculator 21 includes a reference memory access unit 14, a reference memory access unit 15,
_i cumulative value table memory 16, $ d _i cumulative value table memory 18, $ d
_i ² reads data from the cumulative value table memory 19 and? r _i d _i calculator 20, and sends the luminance offset B calculation unit 21 and the square error E calculator 23 calculates a contrast C, and outputs it as the contrast 121 . The luminance offset B calculation unit 22 includes a reference source memory access unit 14,
The data is read from the reference destination memory access unit 15 and the contrast C calculation unit 21, the luminance offset B is calculated and sent to the square error E calculation unit 23, and is output as the luminance offset 123. The square error E calculator 23 calculates
.Sigma.r _i ² cumulative value memory 17, [Sigma] d _i ² cumulative value table memory 19,
.Sigma.r _i d _i calculator 20 reads out the data from the contrast C calculator 21 and luminance offset B calculation unit 22, 2
The squared error E is calculated and sent to the least squared error Emin determination unit 24. The least square error Emin determination unit 24 determines the domain block to which the smallest square error E has been given, and outputs a positional relationship including rotation with the range block as an affine map 120.

[0132] Here, the Σr _i d _i of equation (7),
The calculation method using the above accumulated value table cannot be used. This is because,? R _i d _i in the combination of the domain blocks and range blocks are numerous (several tens of thousand millions or more kinds of) Yes, it is because it is impractical to make a cumulative value table for all the combinations . Therefore, only the? R _i d _i becomes necessary to calculate it, this calculation is the most time-consuming calculations in the approximation image search.

Therefore, it is conceivable to speed up the calculation by using the product-sum operation of the MMX technology as shown in FIG. 16 for speeding up. Note that the example of FIG. 16 illustrates an example in which the product sum operation of each pixel of the reference source partial image 102 (range block 106) and the reference destination partial image 103 (domain block 107) is performed.
The size of the (range block 106) and the reference destination partial image 103 (domain block 107) is 4 × 4 pixels.

That is, in FIG. 16, the four pixels p _{0 to} p ₃ included in the top line of the reference source partial image 102 (106) can be read into the register 108 by one instruction. previous partial image 103 four pixels q ₀ to q ₃ included in the top line of (107) can be loaded into the register 109 with a single instruction. According to the MMX technology, the product-sum operation of each pixel held in each of the registers 108 and 109, that is, each of the operation units 132 and 13
3 to calculate p ₀ × q ₀ + p ₁ × q ₁ and p ₂ × q ₂ + p ₃ × q ₃ and store each value of the operation result in the register 134 can be performed by one instruction.

However, also in this case, when the reference destination partial image (domain block) has a rotational movement relationship, for example, FIG.
The reference partial image 103 (107) of No. 6 is rotated 90 degrees clockwise.
When rotated by degrees, the four pixels q _{0 to} q _{3 at} the left end of the reference destination partial image 103 (107) cannot be read into the register 109 by one instruction, as shown in FIG.

Therefore, in the fifth embodiment of the present invention, as described with reference to FIG.
By rotating the domain image in advance,
Access to the image memory is speeded up.

FIG. 18 shows this state as shown in FIG. That is, in FIG. 18, four pixels p ₀ ~p ₃ included in the top line of the reference original partial image 102 (range block 106) can be loaded into the register 108 with a single instruction. On the other hand, the reference destination partial image 103
Leftmost four pixels q ₀ to q when the (domain block 107) is rotated 90 degrees _{clockwise. 3,} 4 pixels q ₀ to q ₃ horizontal row on the line of the top in the domain block 317
Becomes Therefore, this domain block 317 4
The pixels q _{0 to} q ₃ can be read into the register 109 by one instruction. Thus, in the present embodiment, the product-sum operation of each pixel held in each of the registers 108 and 109, that is, p ₀ × q ₀ + p ₁ × by the operation units 132 and 133.
Calculation of q ₁ , p ₂ × q ₂ + p ₃ × q ₃ and storing each value of the operation result in the register 134 can be executed by one instruction.

Next, FIG. 19 shows a sixth embodiment of the present invention in which the same functions as those of the second embodiment are applied to image coding using fractal coding. 1 shows a schematic configuration of an approximate image search device. In FIG. 19, components having the same functions as those in the above-mentioned respective drawings are denoted by the same reference numerals.

The approximate image search device according to the sixth embodiment shown in FIG. 19 includes a range image memory unit 3 for storing image data of a range image 104, and a reduced image generation unit for generating a domain image 105 from the range image 104. 13, a domain image memory unit 4 for storing image data of the domain image 105, an image rotating unit 6 capable of rotating the range image 104 at a plurality of rotation angles, and a rotated image for storing image data of the rotated range image 414. The memory unit 7, the image data of the range block 106 read from the range image memory unit 3, and the range block 41 from the rotated image memory unit 7.
7, an image selecting unit 10 for selecting any of the image data;
Domain block 107 from domain image memory unit 4
And an image comparing unit 51 for comparing the image data with the range block selected by the image selecting unit 10.

Hereinafter, the operation of the approximate image searching apparatus of the sixth embodiment shown in FIG. 19 will be described. This sixth
In the embodiment, when the approximate image is searched, the range image 104 is set to 0 degree (no rotation), 90 degrees, 180 degrees, and 270 degrees.
A description will be given of a case where the rotation is performed in four different degrees.

In FIG. 19, range image memory unit 3
Stores the image data of the range image 104 that is the input reference source image. The reduced image generation unit 13 generates the domain image 105 by reducing the range image 104 by half. The domain image memory unit 4 includes the reduced image generation unit 1
The domain image 105 generated in step 3 is stored. The image rotating unit 6 generates a range image 414 obtained by rotating the range image 104, and stores the image data in the rotated image memory unit 7.
Expand to Further, the image selecting unit 10 selectively extracts either the image data of the range block 106 supplied from the range image memory unit 3 or the image data of the range block 417 supplied from the rotated image memory unit 7 and performs image comparison. It serves to send to the unit 51.

Here, the image selecting section 10 retrieves each pixel data (image data with a rotation angle of 0 degree) of the range block 106 from the range image memory section 3 and
Send to At this time, the image comparison unit 51 extracts each pixel data of the domain block 107 from the domain image memory unit 4 and obtains a square error with each pixel data of the range block 106 sent from the image selection unit 10. This operation is repeated for each of the plurality of domain blocks 107 included in the domain image 105, and a domain block that gives the smallest square error (least square error) to the range block 106 is obtained.

Next, the image rotating section 6 outputs the range image 104
Is rotated by 90 degrees to create a range image 414, and the image data is expanded in the rotated image memory unit 7. Further, the image selection unit 10 extracts each pixel data of the range block 417 from the rotated image memory unit 7 and sends it to the image comparison unit 51. At this time, the image comparing unit 51 compares each pixel data of the domain block 107 from the domain image memory unit 4 with:
A square error with each pixel data of the range block 417 sent from the image selection unit 10 is obtained. This operation is performed by a plurality of domain blocks 107 included in the domain image 105.
Is repeated, and a domain block that gives the smallest square error (least square error) to the range block 417 is obtained. The above operation is performed at a rotation angle of 180 degrees, 27
The same applies to the case of 0 degrees.

Thereafter, the image comparing section 51 finds a domain block 107 giving the smallest square error from all the domain blocks 107 described above, and outputs the affine mapping 120, the contrast 121, and the luminance offset 122. .

FIG. 20 is a flowchart showing the above processing.

In the flowchart shown in FIG. 20, the approximate image search device shown in FIG.
At step S1, the image data of the range image 104 is input to and stored in the range image memory unit 3.
As 22, the reduced image generation unit 13 generates a domain image 105 obtained by reducing the range image 104 by half, and inputs the domain image 105 to the domain image memory unit 4 and stores it.

Next, in step S123, the image comparing section 51 extracts each pixel data of the range block 106 from the range image memory section 3. At this time, the image comparison unit 51 initializes a register Emin for holding the minimum value of the square error with a very large value in step S124. In the next step S125,
The image selection unit 10 selects the range block 106 from the range image memory unit 3, and the image comparison unit 51 performs an approximate image search between the range block 106 and the domain block 107 by a least square error calculation.

In step S126, the image rotation unit 6 outputs the range image 1 from the range image memory unit 3.
04 is rotated 90 degrees to generate a range image 414, and the image data is stored in the rotated image memory unit 7. In the next step S127, the image selecting unit 10 selects the range block 417 from the rotated image memory unit 7, and the image comparing unit 51 approximates the domain block 107 and the range block 417 by the least square error calculation. Perform an image search.

Similarly, in step S128, image rotating section 6 rotates range image 104 from range image memory section 3 by 180 degrees to generate range image 414, and stores the image data in rotated image memory section 7. Let it. In the next step S129, the image selecting unit 10 selects the range block 417 from the rotated image memory unit 7, and the image comparing unit 51 calculates the approximate image between the domain block 107 and the range block 417 by the least square error calculation. Perform a search.

Further, in step S130, image rotating section 6 rotates range image 104 from range image memory section 3 by 270 degrees to generate range image 414, and stores the image data in rotated image memory section 7. . In the next step S131, the image selecting unit 10 selects the range block 417 from the rotated image memory unit 7, and the image comparing unit 51 calculates the approximate image between the domain block 107 and the range block 417 by the least square error calculation. Perform a search.

Next, in step S132, the image comparing section 51 selects, from all the domain blocks 107,
The domain block 107 that gives the smallest square error is obtained, and its affine mapping 120 and contrast 12
1. Output the luminance offset 122.

Thereafter, in step S133, the approximate image searching device determines whether or not the processing has been completed for all the range blocks 106 and 417 in the range image. Is completed, and when it is determined that the process is not completed, the process returns to step S123.

In the sixth embodiment, the same effect as in the fifth embodiment is obtained by rotating the range image instead of rotating the domain image as in the fifth embodiment. be able to.

Next, FIG. 21 shows a seventh embodiment of the present invention in which the same functions as those in the third embodiment are applied to image coding using fractal coding. 1 shows a schematic configuration of an approximate image search device. In FIG. 21, components having the same functions as those in the respective drawings are given the same reference numerals.

The approximate image search device according to the seventh embodiment shown in FIG. 21 includes a range image memory unit 3 for storing image data of a range image 104, and a reduced image generation unit for generating a domain image 105 from the range image 104. 13, a domain image memory unit 4 for storing image data of the domain image 105, an image rotation unit 6 capable of rotating the domain image 105 at a plurality of rotation angles, and a plurality of rotated domain images 3
A plurality of rotated image memory units 7a, 7b, 7c for storing respective image data of 14a, 314b, 314c;
The image data of the domain block 107 read from the domain image memory unit 4 and the rotated image memory units 7a and 7
b, 7c each domain block 317a, 317
An image selecting unit 10 for selecting one of the image data b and 317c, and an image comparing unit 51 for comparing the range block 106 from the range image memory unit 3 with the domain block selected by the image selecting unit 10. ing.

Hereinafter, the operation of the approximate image searching apparatus of the seventh embodiment shown in FIG. 21 will be described. This seventh
In the embodiment, when the approximate image is searched, the domain image 105 is set to 0 degree (no rotation), 90 degrees, 180 degrees.
The description will be made by taking as an example the case where the rotation is performed in four ways of 270 degrees.

In FIG. 21, range image memory unit 3
Stores the input image data of the range image 104. In the reduced image generation unit 13, the range image 104 is
2, the domain image 105 is generated. The domain image memory unit 4 stores the input image data of the domain image 105. The image rotation unit 6 performs the domain image 1
05 is rotated by 90 degrees, 180 degrees, and 270 degrees to generate a plurality of domain images 314a, 314b, and 314c,
The image data is stored in the corresponding rotated image memory units 7
a, 7b, 7c. The image selecting unit 10 also stores the image data of the domain block 107 supplied from the domain image memory unit 4 and the rotated image memory units 7a, 7b,
7c supplied from the respective domain blocks 317a, 31
7b and 317c are selectively extracted and sent to the image comparison unit 51.

That is, the image selecting unit 10 extracts each pixel data (image data with a rotation angle of 0 °) of the domain block 107 from the domain image memory unit 4 and sends it to the image comparing unit 51. Similarly, the image selection unit 10 sends each domain block 317 from the rotation image memory units 7a, 7b, 7c.
The pixel data a, 317b, and 317c are extracted and sent to the image comparison unit 51.

At this time, the image comparison unit 51 determines whether each pixel data of the range block 106 from the range image memory unit 3 and the domain block 10 sent from the image selection unit 10
The square errors with the respective pixel data of 7, 317a, 317b, and 317c are obtained. This operation is called domain image 1
The domain block that gives the smallest square error (least square error) to the range block 106 is obtained by repeating the process for each of the plurality of domain blocks included in the domain block 05.

Thereafter, the image comparing section 51 checks all the domain blocks 107 and 317a, 317b, 3
A domain block that gives the smallest square error is obtained from 17c, and its affine mapping 120, contrast 121, and luminance offset 122 are output.

FIG. 22 is a flowchart showing the above processing.

In the flowchart shown in FIG. 22, the approximate image searching apparatus shown in FIG.
As 41, the image data of the range image 104 is input to the range image memory unit 3 and stored, and at the same time, at step S1
In step S42, the reduced image generation unit 13 generates a domain image 105 obtained by reducing the range image 104 by 、, and inputs the domain image 105 to the domain image memory unit 4 for storage.

Next, in step S143, the image rotating unit 6 generates a domain image 314a obtained by rotating the domain image 105 from the domain image memory unit 4 by 90 degrees, and stores the image data in the rotated image memory unit 7a. Let it. Similarly, in the image rotation unit 6, as step S 144, the domain image 10
5 is generated by rotating the domain image 314b by 180 degrees, and the image data is stored in the rotated image memory unit 7b.
Then, a domain image 314c obtained by rotating the domain image 105 from 270 degrees is generated, and the image data is stored in the rotated image memory unit 7c.

Next, in step S146, the image comparing section 51 extracts each pixel data of the range block 106 from the register image memory section 3. At this time, the image comparison unit 51 initializes the register Emin for holding the minimum value of the square error with a very large value in step S147.

In the next step S148, the image comparing section 51 performs an approximate image search by the least square error calculation between the range block 106 and the domain block 107 as described above. Similarly, in the image comparison unit 51,
In step S149, range blocks 106 and 90
An approximate image search is performed between the domain block 317a rotated by an angle by a least square error calculation, and step S150 is performed.
In step S151, an approximate image search is performed by the least square error calculation between the range block 106 and the domain block 317b rotated 180 degrees, and in step S151, the domain block 3
Approximate image search by least square error calculation is performed between the image and the image 17c.

Next, in step S152, the image comparing section 51 checks all the domain blocks 107 and 317.
a, 317b, and 317c, the domain block 107 or 317a, 31 that gives the smallest square error.
7b and 317c are obtained, and the affine mapping 120, the contrast 121, and the luminance offset 122 are output.

Thereafter, in step S153, the approximate image search device determines whether or not processing has been completed for all range blocks 106 in the register image.
When it is determined that the processing has been completed, the processing of the approximate image search is terminated, and when it is determined that the processing has not been completed, the processing returns to step S153.

In the seventh embodiment, the same effects as those of the above embodiments can be obtained.

Next, FIG. 23 shows an eighth embodiment of the present invention in which the same functions as those of the fourth embodiment are applied to image coding using fractal coding. 1 shows a schematic configuration of an approximate image search device. Note that, in FIG. 23, components having the same functions as those in the above-described drawings are denoted by the same reference numerals.

An approximate image search device according to the eighth embodiment shown in FIG. 23 includes a range image memory unit 3 for storing image data of a range image 104, and a reduced image generation unit for generating a domain image 105 from the range image 104. 13, a domain image memory unit 4 for storing image data of the domain image 105, an image rotating unit 6 capable of rotating the range image 104 at a plurality of rotation angles, and a plurality of rotated range images 414.
a, 414b, and 414c, respectively, a plurality of rotated image memory units 7a, 7b, and 7c, and image data of the domain block 107 read from the domain image memory unit 4 and the rotated image memory units 7a and 7b. , 7
Each range block 417a, 417b, 417 from c.
c, an image selection unit 10 for selecting any of the image data;
Domain block 107 from domain image memory unit 4
And an image comparing section 51 for comparing the range block selected by the image selecting section 10.

The operation of the approximate image searching apparatus according to the eighth embodiment shown in FIG. 23 will be described below. This 8th
In the embodiment, when the approximate image is searched, the range image 104 is set to 0 degree (no rotation), 90 degrees, 180 degrees,
The description will be made by taking as an example a case where the rotation is performed at four types of 270 degrees.

In FIG. 23, range image memory unit 3
Stores the input image data of the range image 104. In the reduced image generation unit 13, the range image 104 is
2, the domain image 105 is generated. The domain image memory unit 4 stores the input image data of the domain image 105. The image rotation unit 6 outputs the range image 10
4 is rotated by 90 degrees, 180 degrees, and 270 degrees to generate a plurality of range images 414a, 414b, and 414c, and stores the image data in the corresponding rotated image memory units 7a,
7b and 7c. The image selection unit 10 also stores the domain partial image 10 supplied from the domain image memory unit 4.
7 and the rotated image memory units 7a, 7b, 7c
Range blocks 417a, 417b,
417c selectively takes out any of the image data and sends it to the image comparison unit 51.

That is, the image selecting section 10 retrieves each pixel data (image data with a rotation angle of 0 degree) of the range block 104 from the range image memory section 3 and sends it to the image comparing section 51. Similarly, the image selection unit 10 outputs each of the range blocks 417a, 417a,
The pixel data of 417b and 417c are extracted and sent to the image comparison unit 51.

At this time, the image comparison unit 51 determines whether each pixel data of the domain block 107 from the domain image memory unit 4 and the range block 1 sent from the image selection unit 10
A square error with each of the pixel data 06, 417a, 417b, and 417c is obtained. This operation is repeated for a plurality of domain blocks 107 included in the domain image 105, and the range blocks 106, 417a, 41
A domain block that gives the smallest square error (least square error) to 7b and 417c is obtained.

After that, the image comparing section 51 finds a domain block giving the smallest square error from all the domain blocks described above, and obtains the affine mapping 12
0, a contrast 121, and a luminance block 122 are output.

FIG. 24 is a flowchart showing the above processing.

In the flowchart shown in FIG. 24, in the approximate image search device shown in FIG.
As 61, the image data of the range image 104 is input to the range image memory unit 3 and stored, and at the same time, at step S1
As 62, the reduced image generation unit 13 generates a domain image 105 obtained by reducing the range image 104 by １ ／, and inputs and stores the domain image 105 in the domain image memory unit 4.

Next, in step S143, the image rotation unit 6 outputs the range image 1 from the range image memory unit 3.
04 is rotated 90 degrees to generate a range image 414a,
The image data is stored in the rotated image memory unit 7a.
Similarly, in the image rotation unit 6, as step S164,
The range image 104 from the range image memory unit 3 is
A range image 414b rotated by degrees is generated, and the image data is stored in the rotated image memory unit 7b.
65, the range image 1 from the range image memory unit 3
A range image 414c is generated by rotating 04 by 270 degrees, and the image data is stored in the rotated image memory unit 7c.

Next, in step S166, the image comparing section 51 sends the range blocks 106, 4 from the range image memory section 3 and the rotated image memory sections 7a, 7b, 7c.
Each pixel data of 17a, 417b, and 417c is extracted. At this time, the image comparison unit 51 determines in step S16
In step 7, a register Emin for holding the minimum value of the square error is initialized with a very large value.

In the next step S168, the image comparison section 51 performs an approximate image search between the range block 106 and the domain block 107 by a least square error calculation. Similarly, in the image comparison unit 51, step S16
9, an approximate image search is performed between the domain block 107 and the range block 417 a rotated by 90 degrees by a least square error calculation. An approximate image search is performed by a square error calculation, and in step S171, the domain block 1
An approximate image search by least square error calculation is performed between 07 and the range block 417c rotated by 270 degrees.

Next, in step S 172, the image comparison section 51 selects, from all the domain blocks 107,
A domain block that gives the smallest square error is obtained, and its affine mapping 120, contrast 121, and luminance offset 122 are output.

Thereafter, in step S173, the approximate image searching device determines whether or not the processing has been completed for all the range blocks 106, 417a, 417b, and 417c of the range image. The process of the approximate image search is completed, and when it is determined that the process is not completed, the process returns to step S166.

In the eighth embodiment, instead of rotating the reference destination image as in the seventh embodiment,
By rotating the reference source image, the same effect as in the seventh embodiment can be obtained.

In each of the above embodiments, the rotation angle of the image rotation unit 6 is assumed to be 90 degrees, 180 degrees, and 270 degrees. Of course, other rotation angles may be used. Further, the number of rotated image memory units in FIGS. 7, 9, 21, and 23 is not limited to three.

[0185]

According to the approximate image searching apparatus and method of the present invention, a partial image extracted from a reference image or a partial image extracted from an image obtained by rotating a reference image at a plurality of rotation angles is used. By selecting and comparing the partial image extracted from the reference source image and the selected partial image, calculating the difference between them, and selecting the partial image closest to the reference original partial image, Even when the reference destination image is not only in parallel movement but also in rotational movement, continuous access to the image memory (memory access efficiency) is possible in the same way as in the case of parallel movement. Thus, it is possible to suppress a decrease in the execution speed of the operation instruction and to significantly speed up the processing.

In the approximate image searching apparatus and method according to the present invention, the partial image extracted from the reference source image,
Alternatively, a partial image extracted from an image obtained by rotating the reference source image at a plurality of rotation angles is selected, the partial image extracted from the reference destination image is compared with the selected partial image, and the difference is calculated. By selecting the partial image that is the closest to the reference source partial image, if the reference source partial image and the reference destination partial image are not only in the translation but also in the rotation, It is possible to continuously access the image memory (to improve the efficiency of memory access) in the same manner as described above, and it is possible to suppress a decrease in the execution speed of the operation instruction and significantly increase the processing speed.

Further, according to the present invention, by holding a plurality of images that have been rotated at a plurality of angles in advance, only one generation of a rotated image is required, and high-speed processing is possible.

Furthermore, according to the present invention, even when a rotational movement of an image is added to a search for a motion vector or a rotational movement is added to a search for an affine mapping in fractal coding, high-speed processing is performed. Is possible.

[Brief description of the drawings]

FIG. 1 is a block circuit diagram illustrating a schematic configuration of an approximate image search device according to a first embodiment of the present invention.

FIG. 2 is a flowchart illustrating a flow of a process performed by the approximate image searching device according to the first embodiment;

FIG. 3 is a flowchart showing a flow of processing when an approximate image search is performed between a reference source partial image and a reference destination image.

FIG. 4 is a diagram used for describing the configuration and operation when the arithmetic processing in the image comparison unit is implemented using MMX technology in the first embodiment.

FIG. 5 is a block circuit diagram illustrating a schematic configuration of an approximate image search device according to a second embodiment of the present invention.

FIG. 6 is a flowchart illustrating a flow of a process performed by the approximate image searching device according to the second embodiment;

FIG. 7 is a block circuit diagram illustrating a schematic configuration of an approximate image search device according to a third embodiment of the present invention.

FIG. 8 is a flowchart illustrating a flow of a process performed by the approximate image searching device according to the third embodiment;

FIG. 9 is a block circuit diagram illustrating a schematic configuration of an approximate image search device according to a fourth embodiment of the present invention.

FIG. 10 is a flowchart illustrating a flow of a process performed by the approximate image searching device according to the fourth embodiment;

FIG. 11 is a block circuit diagram illustrating a schematic configuration of an approximate image search device according to a fifth embodiment of the present invention.

FIG. 12 is a flowchart illustrating a flow of a process performed by the approximate image searching device according to the fifth embodiment;

FIG. 13 is a flowchart illustrating a flow of processing when an approximate image search is performed between a range image and a domain image.

FIG. 14 is a diagram used to explain a cumulative value table stored in advance in an image comparison unit provided in an approximate image search device when fractal coding is used for image coding.

FIG. 15 is a block circuit diagram illustrating a specific configuration example of an image comparison unit provided in the approximate image search device when fractal coding is used for image coding.

FIG. 16 is a diagram used to describe a configuration and an operation in the case where fractal coding is used for image coding and the arithmetic processing in the image comparison unit is realized using MMX technology.

FIG. 17 is a diagram used to explain a problem in a case where fractal coding is used for image coding and a calculation process in the image comparison unit uses MMX technology.

FIG. 18 is a diagram illustrating a configuration and an operation when an arithmetic process in an image comparison unit is realized using MMX technology in a fifth embodiment.

FIG. 19 is a block circuit diagram illustrating a schematic configuration of an approximate image search device according to a sixth embodiment of the present invention.

FIG. 20 is a flowchart illustrating a flow of a process performed by the approximate image searching device according to the sixth embodiment;

FIG. 21 is a block circuit diagram illustrating a schematic configuration of an approximate image search device according to a seventh embodiment of the present invention.

FIG. 22 is a flowchart illustrating a flow of a process performed by the approximate image searching device according to the seventh embodiment.

FIG. 23 is a block circuit diagram illustrating a schematic configuration of an approximate image search device according to an eighth embodiment of the present invention.

FIG. 24 is a flowchart illustrating a flow of a process performed by the approximate image searching device according to the eighth embodiment;

FIG. 25 is a diagram used for describing an approximate image search in the case of image encoding using a motion vector.

FIG. 26 is a diagram used to explain an approximate image search in the case of image coding using fractal coding.

FIG. 27 is a diagram used to describe the case where image information when the image size is m pixels horizontally and n pixels vertically is handled by a computer or the like.

FIG. 28 is a diagram used to explain a case where image information of a one-dimensional array is handled by a computer or the like.

FIG. 29 is a diagram used to explain a case where image information of a two-dimensional array is handled by a computer or the like.

FIG. 30 shows an example of a conventional SIMD operation function, MM
FIG. 11 is a diagram used for describing a configuration and an operation when calculating a difference between each pixel of a reference source partial image and a reference destination partial image using X technology.

FIG. 31 is a diagram used to explain a problem in calculating a difference between each pixel of a reference source partial image and a reference destination partial image using MMX technology.

[Explanation of symbols]

1 reference source image memory unit, 2 reference destination image memory unit,
3 range image memory unit, 4 domain image memory unit, 6 image rotation unit, 7 rotation image memory unit, 1
0 Image selection unit, 11, 51 Image comparison unit, 13
Reduced image generation unit, 14 reference source memory access unit,
15 referenced memory access unit, 16Σr _i accumulated value table memory, 17? R _i ² cumulative value table memory, 18 [Sigma] d
_i cumulative value table memory, 19 Σd _i ² cumulative value table memory,
20? R _i d _i calculation unit, 21 the contrast C calculator,
22 luminance offset B calculator, 23 square error E
Calculation section

Claims (24)

[Claims] 1. An approximate image search device for searching for another partial image similar to a partial region of an image, a reference source image storage means for storing a reference source image, and a reference destination image storage for storing a reference destination image Means, image rotation means for rotating an image at a plurality of angles, rotation image storage means for storing an image obtained by rotating a reference image at a plurality of rotation angles, read from the reference image storage section An image selection unit for selecting a partial image and a partial image read from the rotated image storage unit; and a partial image read from the reference source image storage unit and a partial image selected by the image selection unit. And an image comparing means for calculating the difference. 2. The approximation according to claim 1, wherein said rotation image storage means comprises a plurality of image memories for respectively storing a plurality of images obtained by rotating a reference destination image by a plurality of angles. Image search device. 3. The approximate image search device according to claim 1, wherein the search for the partial image is used for detecting a motion vector. 4. The approximate image search device according to claim 2, wherein the search for the partial image is used for detecting a motion vector. 5. The approximate image search device according to claim 1, wherein the search for the partial image is used for fractal coding. 6. The approximate image search device according to claim 2, wherein the search for the partial image is used for fractal coding. 7. An approximate image search method for searching for another partial image similar to a partial region of an image, comprising: storing a reference source image; storing a reference destination image; The rotated image is stored, and the partial image extracted from the stored reference image and the partial image extracted from the rotated and stored reference image are selected. A partial image extracted from the reference source image and the selected partial image, and calculating a difference therebetween. 8. The approximate image search method according to claim 7, wherein a plurality of images obtained by rotating the reference image at a plurality of angles are stored separately. 9. The approximate image search method according to claim 7, wherein the search for the partial image is used for detecting a motion vector. 10. The approximate image search method according to claim 8, wherein the search for the partial image is used for detecting a motion vector. 11. The approximate image search method according to claim 7, wherein the search for the partial image is used for fractal coding. 12. The approximate image search method according to claim 8, wherein the search for the partial image is used for fractal coding. 13. An approximate image search device for searching for another partial image similar to a partial area of an image, a reference source image storage means for storing a reference source image, and a reference destination image storage for storing a reference destination image. Means, image rotation means for rotating an image at a plurality of angles, rotation image storage means for storing an image obtained by rotating a reference source image at a plurality of rotation angles, and read from the reference source image storage unit An image selection unit for selecting a partial image and a partial image read from the rotated image storage unit; a partial image read from the reference destination image storage unit and a partial image selected by the image selection unit; And an image comparing means for calculating the difference. 14. The approximation according to claim 13, wherein said rotation image storage means comprises a plurality of image memories for respectively storing a plurality of images obtained by rotating the reference source image at a plurality of angles. Image search device. 15. The approximate image search device according to claim 13, wherein the search for the partial image is used for detecting a motion vector. 16. The approximate image search device according to claim 14, wherein the search for the partial image is used for detecting a motion vector. 17. The approximate image search device according to claim 13, wherein the search for the partial image is used for fractal coding. 18. The approximate image search device according to claim 14, wherein the search for the partial image is used for fractal coding. 19. An approximate image search method for searching for another partial image approximated to a partial region of an image, comprising: storing an image of a reference source; storing an image of a reference destination; The rotated image is stored, and the partial image extracted from the stored reference source image and the partial image extracted from the rotated and stored reference original image are selected. A partial image extracted from an image of a reference destination and the selected partial image, and calculating a difference between the partial image and the selected partial image. 20. The method according to claim 19, wherein a plurality of images obtained by rotating the reference source image at a plurality of angles are stored separately. 21. The approximate image search method according to claim 19, wherein the search for the partial image is used for detecting a motion vector. 22. The approximate image search method according to claim 20, wherein the search for the partial image is used for detecting a motion vector. 23. The approximate image search method according to claim 19, wherein the search for the partial image is used for fractal coding. 24. The approximate image search method according to claim 20, wherein the search for the partial image is used for fractal coding.