JP2017126264A - Information processor, information processing method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information processor capable of increasing the speed in a piece of processing to search a corresponding block between images without reducing the accuracy.SOLUTION: The information processor acquires a plurality of image data. A projection operation is performed on each acquired image data to generate a vector in which the dimensionality corresponding to the pixels in the processing object is reduced. The similarity of pixels processing object is calculated using the generated vector.SELECTED DRAWING: Figure 3

Description

  The present invention relates to processing for searching a corresponding local region from a plurality of images.

  Conventionally, there is a technique for estimating a parallax map or a distance map based on a plurality of images (parallax image groups) having different viewpoints. The parallax map or distance map estimated in this way is used for measuring the three-dimensional shape of the subject. In addition, the parallax map or the distance map may be used to change the focus position, depth of field, viewpoint, illumination, and the like by image processing after shooting. Such a technique for changing the focus position, depth of field, viewpoint, illumination, and the like by image processing after shooting is called “computational photography” and is commercialized in some cameras.

  This parallax map or distance map estimation technique can be divided into a method of searching for corresponding feature points between two images and a method of searching for corresponding blocks. The method for searching for feature points is sparse information, and can only be used supplementarily for the applications described above. On the other hand, the method for searching for a block can be applied to the above-described use because it can estimate dense distance information. When using a method for searching for a block, the estimation accuracy greatly depends on the size of the block. For this reason, a method is generally used in which the process of searching for a corresponding block is executed a plurality of times while changing the block size for the input image. Also, the larger the block size, the higher the accuracy of correspondence, while the amount of calculation increases and the processing time increases. Thus, there is a need for a technique for reducing the amount of computation and speeding up the processing for searching for a corresponding block. Patent Document 1 describes that the amount of calculation is reduced by performing resolution conversion of an input image according to a scene.

JP 2001-126065 A

Dimitris Achryoptas "Database-friendly Random Projections" Proceedings of the 20th ACM SIGMOD-SIGACT-SIGART Symposium 2001, pp. 274-281 Robert Calderbank, Stephen Howard, Sina Jafarpour al., "Construction of a Large Class of Deterministic Sensing Matrices That Satisfy a Statistical Isometry Property" IEEE Journal of Selected Topics in Signal Processing, 2010 years, Vol. 4, No. 2, pp. 358-374

  However, in the technique described in Patent Document 1, since the information amount is reduced by resolution conversion, the accuracy of block matching is reduced when resolution conversion is not appropriate. This is because detailed features on the image disappear when the resolution of the image is reduced. For this reason, there is a restriction that the distance range to which each part of the input image belongs must be accurately estimated in advance.

  An information processing apparatus according to the present invention obtains a plurality of image data, performs a projection operation on each image data obtained by the obtaining means, and reduces the number of dimensions corresponding to a pixel to be processed. It has a generation means which generates a vector, and a calculation means which calculates similarity of the pixel for processing using the vector generated by the generation means.

  According to the present invention, it is possible to speed up the process of searching for a corresponding block between images without reducing accuracy.

1 is a block diagram illustrating a configuration of an information processing apparatus that is used in the description of Embodiment 1. FIG. 1 is a block diagram illustrating a configuration of an information processing apparatus that is used in the description of Embodiment 1. FIG. 3 is a flowchart showing a flow of processing by the information processing apparatus used in the description of the first embodiment. It is a figure which shows an example of the similarity used by description of Example 1. FIG. It is a figure which shows an example of the parallax map used by description of Example 1. FIG. 10 is a flowchart illustrating a flow of processing performed by the information processing apparatus, which is used in the description of the second embodiment. It is a figure which shows an example of the similarity degree used by description of Example 2. FIG. It is a figure which shows an example of the parallax map used by description of Example 2. FIG. It is a schematic diagram explaining the processing content used by description of Example 1 and 2. FIG.

<Example 1>
In the first embodiment, an example will be described in which a corresponding block search process is speeded up by performing dimension compression processing on each local region (hereinafter referred to as a block) of image data. First, the configuration of the information processing apparatus according to the first embodiment will be described.

  FIG. 1 is a diagram illustrating an example of the configuration of the information processing apparatus according to the first embodiment. An information processing apparatus 100 (hereinafter referred to as a processing apparatus 100) according to the first embodiment includes a CPU 101, a RAM 102, a ROM 103, a secondary storage device 104, an input interface 105, and an output interface 106. The components of the processing apparatus 100 are connected to each other via a system bus 107. The processing device 100 is connected to the external storage device 108 and the operation unit 110 via the input interface 105, and is connected to the external storage device 108 and the display device 109 via the output interface 106.

  The CPU 101 is a processor that executes a program stored in the ROM 103 using the RAM 102 as a work memory and controls each component of the processing apparatus 100 through the system bus 107. Thereby, various processes described later are executed. The secondary storage device 104 is a storage device that stores various data handled by the processing device 100, and an HDD is used in this embodiment. The CPU 101 can write data to the secondary storage device 104 and read data stored in the secondary storage device 104 via the system bus 107. In addition to the HDD, various storage devices such as an optical disk drive and a flash memory can be used for the secondary storage device 104.

  The input interface 105 is a serial bus interface such as USB or IEEE1394, for example, and input of data, commands, and the like from an external device to the processing device 100 is performed via the input interface 105. The processing device 100 acquires data from the external storage device 108 (for example, a storage medium such as a hard disk, a memory card, a CF card, an SD card, and a USB memory) via the input interface 105. Further, the processing apparatus 100 acquires a user command input using the operation unit 110 via the input interface 105. The operation unit 110 is an input device such as a mouse or a keyboard, and is used to input a user instruction to the processing device 100.

  Similar to the input interface 105, the output interface 106 includes a serial bus interface such as USB or IEEE1394. In addition, for example, a video output terminal such as DVI or HDMI (registered trademark) can be used. Output of data and the like from the processing apparatus 100 to the external apparatus is performed via the output interface 106. The processing device 100 displays the image by outputting the processed image or the like to the display device 109 (various image display devices such as a liquid crystal display) via the output interface 106. Note that the components of the processing apparatus 100 are not limited to the above, and may include other components, but the description thereof is omitted here. In addition, here, an information processing apparatus such as a PC has been described as an example, but a configuration in which a part of the configuration illustrated in FIG. 1 is network-connected may be used.

  Hereinafter, processing performed by the processing apparatus 100 according to the first embodiment will be described with reference to a functional block diagram illustrated in FIG. 2, a flowchart illustrated in FIG. 3, and a schematic diagram illustrated in FIG. As illustrated in FIG. 2, the processing apparatus 100 according to the first embodiment has functions as a data acquisition unit 201, a projection calculation unit 202, a collation unit 203, and a correspondence determination unit 204. In the processing apparatus 100, the CPU 101 reads and executes a control program stored in the ROM 103, thereby realizing the functions of the above-described units. Note that the processing apparatus 100 may be configured to include a dedicated processing circuit corresponding to each component. Hereinafter, the flow of processing performed by each component will be described.

  In step S <b> 301, the data acquisition unit 201 acquires image data to be processed via the input interface 105 or from the secondary storage device 104. The acquired image data is a plurality of image data that respectively correspond to a plurality of images. For example, the plurality of images include images from a plurality of different viewpoints captured by one or a plurality of cameras. As the imaging means, for example, a multi-lens camera in which a plurality of small cameras are juxtaposed may be used, or a plenoptic camera that can simultaneously acquire images of a plurality of viewpoints by incorporating a microlens array may be used. The processing apparatus 100 determines one of the acquired plurality of images as the reference image 901. As the reference image 901, an image for which information such as a parallax map is to be estimated may be selected from the acquired images, and the user may set in advance, or the processing apparatus 100 may automatically determine the reference image 901. Hereinafter, one of the input images other than the standard image is referred to as a reference image 906.

  In step S <b> 302, the data acquisition unit 201 extracts one block 902 corresponding to one pixel (hereinafter referred to as a target pixel) of the reference image 901, and outputs it to the projection calculation unit 202. A block is composed of a plurality of pixels. The block 902 may be a rectangular area centered on the target pixel.

  In step S <b> 303, the projection calculation unit 202 performs a projection calculation on one block 902 of the reference image 901 input from the data acquisition unit 201. As the projection calculation, there is a method of multiplying a vector 903 (column vector) obtained by transforming the block 902 by a horizontally long projection matrix 904 described later. As a result of this projection operation, a lower-dimensional vector (less elements) 905 is generated from the vector 903 corresponding to the block 902. For example, as shown in FIG. 9A, an M-dimensional vector 905 is generated by multiplying a block of N pixels by a matrix of M × N size. Since the purpose of the projection operation is to reduce the data size, it is assumed that M <N.

  As one of the conditions to be satisfied by the above-described projection matrix, there is a Johnson-Lindenstrauss lemma as described in Non-Patent Document 1. According to this lemma, even if the dimension is reduced by multiplying a difference between two arbitrary vectors by a projection matrix of M <N that satisfies a certain condition, the norm (length) is almost completely preserved. Guaranteed. The information for calculating the similarity between two vectors and the norm of the difference between the two vectors are closely related. This is because when calculating the similarity between two vectors, a process for obtaining a difference between the two vectors is performed. Therefore, even if such a projection matrix is multiplied to convert the block into a lower-dimensional vector, information for calculating the similarity of the block is not lost. As described above, in this embodiment, since the size of data used for calculating the similarity can be reduced without losing information, the calculation amount of the corresponding block search can be reduced while maintaining the accuracy.

  A specific example of a projection matrix in which the norm is preserved even if the dimension is reduced by multiplying the difference between two arbitrary vectors by a projection row example is described in section 1.1 of Non-Patent Document 1. Yes. In Non-Patent Document 1, as a specific example of a projection matrix in which the norm is preserved even if the dimension is reduced by multiplying the difference between two arbitrary vectors by a projection row example, each element is independently 1 A binary matrix taking +1 or -1 at random with a probability of / 2 is described. A sparse matrix in which each element independently has + √3 with a probability of 1/6, −√3 with a probability of 1/6, and 0 with a probability of 2/3 may have the same property. Are known. Note that even if a dimension is reduced by multiplying a difference between two arbitrary vectors by a projection row example, a projection matrix that satisfies the condition that the norm is preserved is not limited to these, and from a vector having randomness The generated cyclic matrix may be used. Furthermore, the projection matrix does not necessarily have randomness, and may be a uniquely determined matrix such as a discrete chirp matrix, Delsarte-Goethals code, or BCH code as described in Section 2 of Non-Patent Document 2. . Note that it is preferable for calculation that the projection matrix is the same regardless of the block. Thus, in the present embodiment, M rows and N columns (M <N) determined by any one of a random number, a discrete chirp matrix, a Delsarte-Goethals code, and a BCH code can be used. Further, in this matrix, there can be substantially the same number of the same values with the signs inverted. Furthermore, the random number may be a value rounded to a discrete value.

  In step S304, the data acquisition unit 201 extracts a plurality of blocks 907 corresponding to the pixel at a specific position from the reference image 906. The position of the block 907 may be a position moved from the position of the block 902 to a plurality of predetermined relative positions. The plurality of relative positions are candidate values of parallax, and may be a set of fixed values regardless of the position of the block 902, or may be a set of values different for each position of the block 902.

  In step S305, the projection calculation unit 202 performs a projection calculation on the vector 908 obtained by transforming each of the blocks 907, and generates a plurality of low-dimensional vectors 909, as in step S303. In the present embodiment, an example has been described in which a plurality of blocks corresponding to the block 902 of the reference image are extracted first, and then a projection operation is performed on each of the extracted blocks. However, the present invention is not limited to this. For example, a single block 907 corresponding to the block 902 of the reference image may be extracted to perform a projection operation, and this processing may be repeated by changing the position of the block 907.

  Note that the projection matrix used in the projection transformation in step S305 is the same projection matrix as the projection matrix used in the projection transformation in step S303. That is, in the first set of the block of the base image and the plurality of blocks of the reference image, which are used in the similarity comparison described later, it is necessary to use the same projection matrix, but in the other sets A projection row example different from the projection matrix used in the first group may be used.

  In step S306, the collation unit 203 calculates the similarity between the vector 905 and the vector 909. That is, the similarity between the vector 905 subjected to the projection operation in step S303 and the plurality of vectors 909 subjected to the projection operation in step S305 is calculated. As a method of calculating the similarity, a generally known sum of squared differences, sum of absolute differences, normalized cross-correlation, or the like may be used, but the method is not limited to these. In conventional block matching, the similarity between the pre-projection vector 903 and the vector 908 is calculated, but the amount of calculation increases as the block size increases. On the other hand, in the method described in this embodiment, the calculation amount can be reduced because the similarity between vectors with reduced dimensions is calculated.

  In step S307, the correspondence determination unit 204 estimates the parallax at the position of the block 902 of the reference image 901 based on the similarity calculated in step S306 and the relative position of the block 907 of the reference image to the block 902 of the reference image. . As the simplest method, the distance d between the block 902 and the block 907 having the maximum similarity is used as the parallax value. As another method, the degree of similarity is regarded as a function of the distance between the block 902 and the block 907, a distance having a maximum value is calculated by function fitting, and this is used as a parallax value. Note that the method of estimating the parallax based on the similarity is not limited to these, and any known technique can be used. Further, the parallax value may be converted into a distance value as necessary.

  In step S308, the processing device 100 determines whether processing has been performed on the block 902 corresponding to all the pixels of the target pixel for which the parallax of the reference image 901 is to be calculated. If the process has not been completed, the process returns to step S302 to select the block 902 corresponding to the unprocessed target pixel, and then perform the processes of steps S302 to 307.

  The above is the process performed by the processing apparatus 100 according to the first embodiment. This process is schematically shown in FIG. According to the above process, the number of dimensions of the blocks in the image can be reduced without reducing the amount of information, so that a parallax map can be generated at high speed.

  In order to explain the effect of this embodiment, an example in which the above processing is actually performed on image data will be shown below.

Two stereo images having parallax only in the horizontal direction are set as input image data. The block size is 5 × 5 pixels, and the normalized cross-correlation NCC shown in Expression (1) is calculated for each reference image coordinate (x ₀ , y ₀ ) and estimated parallax d as the similarity to each block set.

Here, F (x, y) is the pixel value at the coordinates (x, y) of the standard image, G (x, y) is the pixel value at the coordinates (x, y) of the reference image, and B is the coordinates (x ₀ , It is a set of coordinates in a block centered on y ₀ ). When the block size is 5 × 5 pixels, B in the case where no projection calculation is performed is a set of 25 sets of coordinates. When performing projection calculation, B is a set of 10 coordinates, the matrix to be multiplied by the vector corresponding to the block is 10 × 25 size, and each element independently takes 0 or 1 with a probability of 1/2. It was. That is, the vector dimension is reduced by 60% by the projection operation.

For each coordinate (x ₀ , y ₀ ), search for d that maximizes the NCC, and let d be the value at the coordinate (x ₀ , y ₀ ) of the parallax map. FIG. 4 shows the correspondence between NCC and d in one coordinate (x ₀ , y ₀ ). FIG. 4A shows a case where the projection calculation is not performed, and FIG. 4B shows a case where the projection calculation is performed. In FIG. 4, Δ represents a point having the maximum NCC. Regardless of the projection operation, the values of d at which the NCC takes the maximum value are the same, and the dCC dependency of the NCC is similar. In addition, FIG. 5 shows a result of performing 5 × 5 pixel median filter processing on the calculated parallax map. FIG. 5A shows a distribution obtained by subtracting FIG. 5A from FIG. 5B when FIG. 5B is a case where projection operation is not performed, FIG. 5B is a case where projection operation is performed, and FIG. It is. Except for positions where it is difficult to estimate parallax in principle, such as near the edge of an image or the contour of an object, the difference between them is almost zero. As described above, even if the dimension of the vector used for calculating the similarity is reduced, the block information is preserved if the projection matrix satisfies a predetermined condition. Therefore, the parallax estimation result is not substantially affected. .

  The above is the processing of the first embodiment. According to the above processing, it is possible to reduce the amount of calculation of the similarity calculation processing without impairing the estimation accuracy of the parallax map.

<Example 2>
In the method of the first embodiment, each block is extracted from the image data, vectorized, and multiplied by a matrix, so that the calculation efficiency is not good. Therefore, in the second embodiment, a plurality of kernels having the same size and generated by the same method as the projection matrix described above are generated. A plurality of projected images are calculated by convolving each of the generated kernels with the standard image and the reference image, and the similarity is calculated using the projected images. Since the two-dimensional convolution operation can be calculated efficiently using the fast Fourier transform, this method can be processed faster than multiplying the matrix for each block.

  In this embodiment, instead of using a low-dimensional vector obtained by projective calculation from a vector corresponding to a block, a vector in which pixel values of the same coordinates of a plurality of projected images obtained as described above are arranged is used for similarity. Except for use in the calculation, the second embodiment is the same as the first embodiment.

  Hereinafter, processing performed by the processing apparatus 100 according to the first embodiment will be described with reference to a functional block diagram illustrated in FIG. 2, a flowchart illustrated in FIG. 6, and a schematic diagram illustrated in FIG. 9B.

  In step S601, as in step S301, the data acquisition unit 201 acquires a plurality of processing target image data corresponding to a plurality of images via the input interface 105 or from the secondary storage device 104. Further, the processing apparatus 100 determines one of the plurality of images as the standard image 901 and the other as the reference image 906.

  In step S <b> 602, the projection calculation unit 202 convolves a plurality of kernels 910 with the reference image 901, generates a projection image group 911 including projection images corresponding to the kernels, and outputs the projection image group 911 to the collation unit 203. Similarly, a plurality of kernels 910 that are the same as those described above are convoluted with respect to the reference image 906, and a projected image group 913 including projected images corresponding to the kernels is generated and output to the matching unit 203. Note that the kernel used for convolution here is a kernel in which the values of elements in the kernel are generated in the same manner as in the projection row example described in the first embodiment. By the processing in step S602, projection image groups 911 and 913 corresponding to the number of kernels are respectively generated. In addition, the projection calculation unit 202 generates a projection image using a plurality of kernels in the same order for the base image 901 and the reference image 906, and outputs the projection image to the matching unit 203 in the order of generation. The collation unit 203 processes the projected image group so that the correspondence between the projected image of the standard image 901 and the projected image of the reference image 906 obtained using the same kernel can be understood. For example, the collation unit acquires a projected image group in the order output from the projection calculation unit 202.

  In step S603, the collation unit 203 selects one pixel position from which the pixel group 912 is extracted from the projected image group 911 of the reference image. That is, the pixel position of the target pixel in the projected image group 911 of the reference image is selected.

  In step S <b> 604, the matching unit 203 extracts a pixel group 912 at the same position as the target pixel of the reference image projection image group 911 from the projection image group 911 and generates a vector 905.

  In step S605, the collation unit 203 extracts a plurality of pixel groups 914 corresponding to a specific position from the projected image group 913 in the projected image group 913 of the reference image, and generates a plurality of vectors 909. Note that the position of the pixel group 914 may be a position moved from the pixel group 914 to a plurality of predetermined relative positions. Thus, the relative position between the pixel group 912 and the pixel group 914 may be the same as the relative position described in step S304.

  As described above, the collation unit 203 processes the projected image group so that the correspondence between the projected image of the base image 901 and the projected image of the reference image 906 obtained using the same kernel can be understood. Therefore, each element in the generated vector 905 and vector 909 is an element based on a projection image generated using the same kernel.

  In step S606, the collation unit 203 calculates the similarity between the vector 905 and the vector 909. As a result of performing the processing in this way, the amount of calculation is reduced compared to the case where a block having the same size as the kernel is directly used for calculating the similarity.

  In step S607, as in step S307, the correspondence determining unit 204 selects the selected pixel position of the reference image based on the similarity calculated in step S606 and the distance between the pixel positions of the pixel group 912 and the pixel group 914. The parallax at is estimated. As the simplest method, the distance between the pixel positions of the pixel group 912 and the pixel group 914 having the maximum similarity is used as the parallax value. As another method, the similarity is regarded as a function of the distance between the pixel positions of the pixel group 912 and the pixel group 914, the distance between the pixel positions having the maximum value is calculated by function fitting, and this is used as the parallax value.

  In step S608, it is determined whether processing has been performed on all pixels to be processed for which the parallax of the reference image is to be calculated. If the process has not been completed, the process returns to step S603, and after selecting a pixel position corresponding to an unprocessed pixel, the processes of steps S603 to S607 are performed.

  The above is the process performed by the processing apparatus 100 according to the second embodiment. According to the above processing, it is possible to reduce the necessary information for the blocks in the image without losing them, so that the parallax map can be generated at high speed.

  The order of the calculation amount is O (Nn) when the projection calculation is not used when NCC is used for the similarity and the input image has n pixels. On the other hand, in the case of using the projection operation of the present embodiment, O (nlogn) + O (Mn) is obtained because two operations of convolution using fast Fourier transform and NCC calculation using a reduced block are performed. When these are compared, the ratio is approximately N to M, and it can be seen that the smaller M is, the faster the method of this embodiment is. That is, the calculation of similarity using a vector is faster as the number of dimensions (M) is smaller as in this embodiment. Note that the number of dimensions of the vector in this embodiment corresponds to the number of kernels. Therefore, the total number of kernels used in this embodiment is assumed to be smaller than the number of pixels included in the kernel.

  In order to explain the effect of this embodiment, an example in which the above processing is actually performed on image data will be shown below.

FIG. 7 shows the correspondence between NCC and d at the same coordinate (x ₀ , y ₀ ) as in FIG. Compared with FIG. 4A, the values of d at which the NCC takes the maximum value match, and the dCC dependency of the NCC is similar. In addition, FIG. 8 shows the result of performing 5 × 5 pixel median filter processing on the calculated parallax map. FIG. 8A shows a distribution in which no projection calculation is performed, and FIG. 8B shows a distribution obtained by subtracting FIG. 5A from FIG. 8A. Similar to the first embodiment, except for the positions that are difficult to estimate, the difference between them is almost zero. As an example, the calculation time required to calculate the parallax map is 6.4 seconds in FIG. 5A and 3.7 seconds in FIG. 8A.

  The above is the processing of the second embodiment. According to the above processing, it is possible to realize high-speed processing without impairing the estimation accuracy of the parallax map.

<Other embodiments>
In the embodiment described above, the process of searching for a corresponding local region (partial region) of two captured images with different shooting conditions has been described as an example. However, the processing is not necessarily limited to a captured image, and any processing that searches for a corresponding local region between two images may be used, and an image obtained by how the processing target image may be obtained. Further, here, the processing for searching for a local region between two images has been described as an example, but the number of images to be processed is not limited to two, and a local region between a plurality of images is searched. Such processing may be used.

  The present invention supplies a program that realizes one or more functions of the above-described embodiments to a system or apparatus via a network or a storage medium, and one or more processors in a computer of the system or apparatus read and execute the program This process can be realized. It can also be realized by a circuit (for example, ASIC) that realizes one or more functions.

201 Data Acquisition Unit 202 Projection Operation Unit 203 Collation Unit 204 Correspondence Determination Unit

Claims (12)

Obtaining means for obtaining a plurality of image data;
Projection calculation is performed on each image data acquired by the acquisition unit, and a generation unit that generates a vector with a reduced number of dimensions corresponding to a pixel to be processed;
An information processing apparatus comprising: a calculation unit that calculates the similarity of the pixel to be processed using the vector generated by the generation unit.   The generation means includes a projection matrix of M rows and N columns (M <N) determined by any one of a random number, a discrete chirp matrix, a Delsarte-Goethals code, and a BCH code, and each image indicated by the plurality of image data. The information processing apparatus according to claim 1, wherein a vector with a reduced number of dimensions is generated by obtaining a product of a vector corresponding to a pixel value of a partial region composed of a plurality of pixels.   The generating means performs a convolution operation between a plurality of kernels each including a value determined by any one of a random number, a discrete chirp matrix, a Delsarte-Goethals code, and a BCH code and each image indicated by the plurality of image data. 2. The information processing apparatus according to claim 1, wherein a projection image is generated using a vector, and a vector corresponding to a pixel value of a pixel group at a predetermined position of the projection image is generated as the vector with the reduced number of dimensions. .   The generating means is a pixel group at a predetermined position of a first projected image group obtained by performing a convolution operation of the plurality of kernels on a first image among a plurality of images indicated by a plurality of image data. A second projected image group obtained by performing a convolution operation of the plurality of kernels on a value of each element included in the vector corresponding to the pixel value of the second image and a second image different from the first image The vector so that the value of each element included in the vector corresponding to the pixel value of the pixel group at the predetermined position of the predetermined value becomes a value corresponding to the projected image obtained by convolving the same kernel with each image. The information processing apparatus according to claim 3, wherein:   The information processing apparatus according to claim 3, wherein a total number of the kernels is smaller than a number of elements included in the kernel.   The information processing apparatus according to claim 2, wherein in the projection matrix, there are approximately the same number of the same values with the signs inverted.   6. The information processing apparatus according to claim 3, wherein in the kernel, there are substantially the same number of the same values with inverted signs.   The information processing apparatus according to claim 2, wherein the matrix is a cyclic matrix.   The information processing apparatus according to claim 2, wherein the random number is a value rounded to a discrete value.   The information processing apparatus according to claim 1, further comprising a calculation unit that calculates parallax or a distance between the subject and the camera using the similarity. An acquisition step of acquiring a plurality of image data;
A projecting operation is performed on each image data acquired in the acquisition step, and a generation step for generating a vector with a reduced number of dimensions corresponding to the pixel to be processed is provided.
A calculation step of calculating a similarity of the pixel to be processed using the vector generated in the generation step.   The program for functioning a computer as each means of the information processing apparatus as described in any one of Claims 1-10.