JP6460332B2 - Feature value generation unit, collation device, feature value generation method, and feature value generation program - Google Patents

Description

  The present invention relates to a technique for generating a feature amount used for object matching by processing an object image in which the object is captured.

  Conventionally, image processing techniques for detecting a specific object (search target object) from objects (search target objects) captured in a frame image captured by a camera have been used in various fields. For example, in criminal investigations, etc., it is used for searching whether or not a person (suspect) related to a crime is captured with respect to a moving image captured by a security camera installed in various places. Further, in the production line, an image obtained by imaging an article flowing through the production line is processed and used to detect different kinds of articles and defective articles (articles that are not search target objects).

  Whether or not the search target object is a search target object is determined based on the feature amount of the search target object generated from the image in which the search target object is captured and the search target generated from the image in which the search target object is captured. The feature amount of the target object is collated and the degree of similarity is used. Therefore, in order to improve the accuracy of object collation, it is important to generate an object feature value suitable for collation from an image obtained by capturing the object.

  For example, Patent Documents 1 and 2 disclose a method of generating a feature amount including information on the color and shape of an object from an image obtained by capturing the object. Specifically, the feature amount of the object is generated by the binarized gray value of the pixel of the image in which the object is captured and the local autocorrelation of the quantized color vector.

JP 2006-260310 A JP 2010-26712 A

  Recently, suspects may be searched over a wide area as a countermeasure against terrorism incidents, violent crimes, and the like. In this case, the captured images of a number of security cameras with different imaging areas are confirmed. That is, a suspect is collated between security cameras. Between different security cameras, there are differences in performance such as resolution and imaging environments such as the angle to the imaging area and surrounding brightness. The suspect will be collated. For this reason, there is a demand for development of a technique for generating feature quantities of objects that are robust against differences in camera performance and imaging environment from captured images of objects.

  An object of the present invention is to provide a technique for generating object feature quantities that are robust against differences in camera performance and imaging environment from captured images of an object, and improving the accuracy of object collation. To do.

  In order to achieve the above object, the feature value generation unit of the present invention is configured as follows.

  The feature value generation unit generates a feature value of the global region set for the object image in which the object is captured.

The object image is an image obtained by cutting out an area where the object is imaged from the frame image where the object is imaged and converting the extracted area into a predetermined number of pixels . Thereby, the size of the object image for generating the feature amount of the global region becomes uniform. The entire object image may be a single global area, or a plurality of global areas may be set for the object image. In the case of setting a plurality of global areas for the object image, to each global area may be set so as not to overlap the other global areas, be set to overlap the other global areas Good. The feature amount of the object captured in the object image is the feature amount of each global area set for the object image.

  The pixel feature vector extraction unit extracts a pixel feature vector for each pixel of the object image in which the object is captured. The element of the pixel feature vector includes an element related to a luminance value, an element related to an edge, an element related to a hue, and the like.

  The first autocorrelation matrix calculation unit calculates an autocorrelation matrix of pixel feature vectors of pixels located in the local region for each of the plurality of local regions set in the global region. The autocorrelation matrix includes information on the average of the pixel feature vectors (the autocorrelation matrix is a covariance matrix when the average is zero).

The local region vector generation unit generates a local region vector from the autocorrelation matrix of the pixel feature vector calculated by the first autocorrelation matrix calculation unit for each local region. For example, in order to appropriately calculate the distance between autocorrelation matrices, after performing logarithmic transformation (matrix logarithmic transformation) defined for the matrix, the autocorrelation matrix of the image feature vector is rearranged into a vector, Generate a region vector.

  The second autocorrelation matrix calculation unit calculates an autocorrelation matrix of a local region vector of a local region located in the global region. Then, the global region vector generation unit generates a global region vector from the autocorrelation matrix of the local region vector calculated by the second autocorrelation matrix calculation unit as a feature amount of the global region. The global region vector generation unit also generates a global region vector, for example, by performing matrix logarithmic transformation on the autocorrelation matrix of the local region vector and rearranging the vectors into vectors.

  In this way, the pixel feature distribution (autocorrelation matrix) in the local region is once calculated as the local feature, and then the local feature distribution (autocorrelation matrix) in the global region is calculated as the global feature. Calculate as a quantity. The feature amount of the global region is a feature distribution of a local region located in the global region. Therefore, even if an object having a similar distribution amount of a feature amount (pixel feature vector) such as a color or gradient in the global region is an object having a different pixel distribution in the local region, the different global region Feature vectors can be generated, and the discriminability of the object is improved. Also, the pixel feature distribution in the local region of the image is robust against differences in camera performance and imaging environment. Therefore, the feature amount of the global region obtained by integrating the pixel feature distribution in the local region of the image is also robust against differences in camera performance and imaging environment.

  Further, the collation device according to the present invention uses the global area vector generated for the search target object by the feature quantity generation unit and the global area vector generated for the search target object, and the search target object is the search target object. Determine if it exists. As described above, the feature quantity in the global area is robust against differences in camera performance and imaging environment, and therefore the object collation accuracy can be improved.

  According to the present invention, it is possible to generate a feature quantity of an object that is robust against differences in camera performance and imaging environment from a captured image of the object.

It is a block diagram which shows the structure of the principal part of a collation apparatus. It is a schematic functional block diagram which shows the function structure of an image process part. It is a flowchart which shows a feature-value production | generation process. It is a figure which shows a process target frame image. It is a figure which shows the object area | region set on the process target frame image. It is a figure which shows a normalized image. It is a figure which shows the global region set with respect to a normalized image. It is a figure which shows the reference pixel set with respect to a normalized image. It is a flowchart which shows a collation process.

  Hereinafter, a collation device according to an embodiment of the present invention will be described.

  FIG. 1 is a block diagram showing a configuration of a main part of a collation device according to this example. The collation apparatus 1 includes an image input unit 2, an image processing unit 3, a collation unit 4, an output unit 5, and a feature amount storage unit 6.

  The collation apparatus 1 according to this example performs a feature amount generation process for generating a feature amount of a subject who is a person captured in a frame image. The collation device 1 collates the feature quantity of the target person (search target person) generated by the feature quantity generation process with the feature quantity of another target person (search target person) generated by the feature quantity generation process. A collation process for determining whether the search target person is a search target person is performed.

  The image input unit 2 receives a still image or a moving image in which the target person is captured from the connected external device. The external device may be any device as long as the device has a function of inputting a still image or a moving image to the image input unit 2. For example, the external device may be a security camera that captures an image of the monitoring area, or may be a playback device that plays back a moving image file related to a moving image in which the security camera images the monitoring area. It may be a digital still camera that stores still images.

  The image processing unit 3 performs a feature amount generation process for generating a feature amount of the target person who is a person captured in the frame image input to the image input unit 2. The image processing unit 3 executes the feature value generation method according to the present invention. Further, the feature value generation program according to the present invention causes an image processing processor (computer) included in the image processing unit 3 to execute each step. Details of the feature amount generation processing executed by the image processing unit 3 will be described later.

  The collation unit 4 collates the feature quantity of the target person (search target person) with the feature quantity of another target person (search target person), and determines whether or not the search target person is a search target person. Perform verification processing. Details of the matching process executed by the matching unit 4 will be described later.

  The output unit 5 outputs the result of the collation process in the collation unit to a higher-level device or the like.

  The feature amount storage unit 6 identifies, for each target person for which the image processing unit 3 has generated the feature amount, an ID for identifying the target person, the feature amount generated for the target person, and the imaging date and time of the frame image used to generate the feature amount. The collation data in which the imaging location of the frame image used for generating the feature amount is associated is stored. The feature amount storage unit 6 stores the verification data of the target person in the feature amount storage unit 6 for each frame image in which the feature amount of the target person is generated. Therefore, even if the person is the same person, if the frame images in which the feature values are generated are different, matching data with different IDs is stored in the feature value storage unit 6.

  FIG. 2 is a schematic functional block diagram showing a functional configuration of the image processing unit. The image processing unit 3 includes an object detection function unit 31, an object region extraction function unit 32, an image size normalization function unit 33, a pixel feature vector map generation function unit 34, a region setting function unit 35, a local region A vector generation function unit 36, a global area vector generation function unit 37, and a feature amount generation function unit 38 are included.

  The object detection function unit 31 detects a person (subject) captured in the frame image input to the image input unit 2. In addition, the object detection function unit 31 has a function of selecting a processing target frame image from frame images in which the target person is captured.

  The object area extraction function unit 32 extracts a rectangular area surrounding the target detected by the object detection function unit 31 as an object area on the processing target frame image selected by the object detection function unit 31.

  The image size normalization function unit 33 converts the image of the object area extracted by the object area extraction function unit 32 into an image having a predetermined number of pixels. The image size normalization function unit 33 converts the image of the object region cut out by the object region extraction function unit 32 into an image having a predetermined number of pixels by a known nearest neighbor method, bilinear method, bicubic method, or the like. . The image of the object area converted by the image size normalization function unit 33 to a predetermined number of pixels is referred to herein as a normalized image.

  The object area extraction function unit 32 and the image size normalization function unit 33 correspond to the object image generation unit referred to in the present invention.

  The pixel feature vector map generation function unit 34 generates a pixel feature vector for each pixel of the normalized image. The pixel feature vector is a d-dimensional vector (d> 1). The pixel feature vector map generation function unit 34 corresponds to the pixel feature vector extraction unit referred to in the present invention.

  The area setting function unit 35 sets a local area and a global area for the normalized image. The local area is an area composed of a plurality of pixels. The global region is a region where a plurality of local regions are located. That is, the global region is larger than the local region. In this example, a plurality of global regions are set for the normalized image. The plurality of global areas may be set so as not to overlap with other global areas, or may be set so as to overlap with other global areas. The local area may be set so as not to overlap with other local areas, or may be set so as to overlap with other local areas.

  For each local region set by the region setting function unit 35, the local region vector generation function unit 36 calculates an autocorrelation matrix of pixel feature vectors of pixels located in the local region. An autocorrelation matrix of pixel feature vectors of pixels located in the local region corresponds to the first autocorrelation matrix referred to in the present invention.

  Further, the local region vector generation function unit 36 generates, for each local region, a local region vector obtained by vectorizing the autocorrelation matrix of the pixel feature vector of the pixel calculated for the local region. The local region vector is a feature amount of the local region. The local region vector generation function unit 36 corresponds to the first autocorrelation matrix calculation unit and the local region vector generation unit referred to in the present invention.

  For each global region set by the region setting function unit 35, the global region vector generation function unit 37 calculates an autocorrelation matrix of local region vectors of local regions located in the global region. The autocorrelation matrix of the local region vector of the local region located in the global region corresponds to the second autocorrelation matrix referred to in the present invention.

  The global area vector generation function unit 37 generates a global area vector obtained by vectorizing the autocorrelation matrix of the local area vector of the local area calculated for each global area. The global area vector is a feature quantity of the global area. The global region vector generation function unit 37 corresponds to the second autocorrelation matrix calculation unit and the global region vector generation unit referred to in the present invention.

  The feature quantity generation function unit 38 uses the global area vector of each global area as the feature quantity of the subject. That is, the feature amount of the subject is a plurality of global region vectors (the number of global regions set for the normalized image by the region setting function unit 35).

  Hereinafter, the operation of the verification device 1 according to this example will be described.

  First, the feature value generation process in the verification device 1 will be described. FIG. 3 is a flowchart showing the feature amount generation processing. The image processing unit 3 executes this feature amount generation process. Here, the frame image input to the image input unit 2 is a color image. In addition, information such as the imaging date and time and the imaging location of the frame image input to the image input unit 2 is input to the verification device 1.

  The image processing unit 3 selects a processing target frame image from the frame images input to the image input unit 2 (s1). In s1, the object detection function unit 31 determines whether or not the target person (target person who generates the feature amount) is captured in the frame image input to the image input unit 2 by a known method such as pattern matching. . The object detection function unit 31 selects a frame image in which the subject is captured as a processing target frame image.

  Note that if the image input to the image input unit 2 is a moving image, the subject may be captured in a plurality of temporally continuous frame images. When the subject is captured in a plurality of temporally continuous frame images, the object detection function unit 31 is optimal for generating the feature amount in the subject's imaging state among these frame images. A certain frame image is selected as a processing target frame image. In this example, a frame image whose size of the subject on the frame image is closest to the normalized size is selected as the processing target frame image.

  FIG. 4 is a diagram illustrating the processing target frame image selected in s1. The object area cutout function unit 32 sets a rectangular area surrounding the subject imaged in the processing target frame image selected in s1 as an object area, and cuts out the object area (s2). FIG. 5 is a diagram showing an object area set on the processing target frame image.

  The image size normalization function unit 33 converts the image of the object region cut out in s2 into a normalized image having a predetermined number of pixels (n (horizontal) × m (vertical) pixel image). (S3). In s3, the image of the object region cut out in s2 is converted into a normalized image by a known nearest neighbor method, bilinear method, bicubic method, or the like. FIG. 6 shows a normalized image. In this example, the normalized image is 48 (= n) × 128 (= m) pixels.

  As described above, since the frame image whose target person size on the frame image is closest to the normalized size is selected as the processing target frame image, the influence of the normalized image generation processing relating to s3 can be suppressed.

  The pixel feature vector map generation function unit 34 creates, for each pixel of the normalized image, a pixel feature vector map in which the pixel feature vector of the pixel is registered (s4). In this example, the pixel feature vector x (r) of an arbitrary pixel r in the normalized image is a 6-dimensional vector represented by [Equation 1]. r indicates the position of the pixel in the normalized image.

In [Expression 1], R (r) is the red luminance value of the pixel, G (r) is the green luminance value of the pixel, B (r) is the blue luminance value of the pixel, and I _x is The horizontal edge strength, I _y, is the vertical edge strength. || is an absolute value. The pixel feature vector x (r) shown in [Expression 1] includes information regarding the color and pattern of the object.

  The pixel feature vector is not limited to the 6-dimensional vector shown in [Equation 1], but may be a 9-dimensional vector including elements related to hue (H), saturation (S), and brightness (V). Alternatively, a 6-dimensional vector in which elements relating to RGB luminance values are replaced with elements relating to hue (H), saturation (S), and brightness (V) may be used. Furthermore, when the image input to the image input unit 2 is not a color image (in the case of a monochrome image), a four-dimensional vector in which elements relating to RGB luminance values are replaced with luminance values of the pixels is used. Also good.

  Further, the strength of the edge in the horizontal direction and the strength of the edge in the vertical direction can be detected by using a known Prewitt filter, Sobel filter, or the like.

  The area setting function unit 35 performs area setting processing for setting the global area and the local area for the normalized image generated in s3 (s5). The process related to s5 may be performed before the process related to s4. FIG. 7 is a diagram illustrating a global region set for a normalized image. In this example, as shown in FIG. 7, seven global regions A to G are set for the normalized image. The sizes of the global areas A to G are the same. In this example, the size of the global area is 48 × 32 pixels (1536 pixels). The area setting function unit 35 equally divides the normalized image into four in the vertical direction (vertical direction shown in FIG. 7), and sets the global area A, the global area C, the global area E, and the global area G in order from the top. . Further, the area setting function unit 35 has a global area B in which the upper half overlaps the global area A, the lower half overlaps the global area C, the upper half overlaps the global area C, and the lower half overlaps the global area E. And a global region F in which the upper half overlaps the global region E and the lower half overlaps the global region G is set.

  Note that the number of global regions set for the normalized image is not limited to seven shown in FIG. 7 and may be any number. Also, one of the global areas may be set so as to overlap with another global area with respect to the normalized image, or may be set so as not to overlap.

  The region setting function unit 35 sets a plurality of local regions by setting a plurality of reference pixels p for the normalized image. The local region is an α × β pixel region centered on the reference pixel p for each set reference pixel p. In this example, the local region is a region of 7 (= α) × 7 (= β) pixels (49 pixels) centered on the reference pixel p. As shown in FIG. 8, the region setting function unit 35 sets reference pixels p at intervals of two pixels in the horizontal direction and the vertical direction with respect to the normalized image. In FIG. 8, each pixel indicated by hatching is a reference pixel p. FIG. 8 illustrates one local region set for one reference pixel p.

  Note that the region setting function unit 35 may set all the pixels of the normalized image as reference pixels p, or may set the reference pixels p at intervals of three pixels in the horizontal direction and the vertical direction. Further, the size of the local region is not limited to 7 × 7 pixels, and may be set to 3 × 3 pixels, 4 × 4 pixels, 11 × 11 pixels, or the like. Furthermore, in this example, the local region is set around the reference pixel p. However, the reference pixel p may not be the center, and the reference pixel p may be a pixel at the upper right corner of the local region. The pixel in the upper left corner may be used. That is, the local region may be any region set as long as it is a region that is set based on the reference pixel p.

  The global region setting condition for the normalized image, the reference pixel p setting condition for the normalized image, the local region setting condition based on the reference pixel p, and the like are set in advance.

  For each reference pixel p set in s5, the local region vector generation function unit 36 creates an autocorrelation station matrix map in which the autocorrelation matrix Σ (p) of the local region set for the reference pixel p is registered. (S6). The autocorrelation matrix map generated in s6 associates the reference pixel p with the local area autocorrelation station matrix Σ (p).

  The autocorrelation matrix Σ (p) of the local region set for the reference pixel p is

Is calculated by In [Expression 2], NL (p) is the number of pixels in the local area of the reference pixel p (49 pixels in this example), and LocalArea (p) is the local area set for the reference pixel p. Indicates.

  The local area vector generation function unit 36 calculates an autocorrelation matrix Σ (p) for each local area LocalArea (p) set for the reference pixel p of the normalized image. At this time, since the autocorrelation matrix of each local region is a symmetric matrix, one of the overlapping elements is calculated, and the other calculation result is used for the other. In this example, since the pixel feature vector x (r) is 6-dimensional, the autocorrelation matrix of the pixel feature vector x (r) is 6 rows and 6 columns. In addition, there are 15 sets of overlapping elements. That is, in this example, in the calculation of the autocorrelation matrix of one local region, the calculation for 15 elements can be omitted.

  The local region vector generation function unit 36 performs matrix logarithmic transformation on the autocorrelation matrix Σ (p) for each local region LocalArea (p) (s7). Since the autocorrelation matrix is a positive definite target matrix, the space formed by this matrix is not a Euclidean space but a Riemannian manifold. Matrix logarithmic transformation is equivalent to projecting a positive definite target matrix to the tangent space of this manifold, and after doing this, the matrix is vectorized, so that the operation considering the manifold structure of the autocorrelation matrix can be performed. It can be realized by a normal Euclidean space calculation.

  In s7, the autocorrelation matrix Σ (p) of each local area LocalArea (p) is

に示すように特異値分解する。

Singular value decomposition as shown in

  Also, the matrix logarithmic transformation

と定義する。この行列対数変換では、固有値λが０であるとき、行列対数変換した値が−∞になるのを防止するため、特異値λにεを加算するようにしている。この例では、ε＝１とした。

It is defined as In this matrix logarithmic transformation, when the eigenvalue λ is 0, ε is added to the singular value λ to prevent the matrix logarithmic transformed value from becoming −∞. In this example, ε = 1.

  Further, the local region vector generation function unit 36 considers the overlap of elements in the matrix Y (p) = log (Σ (p)) subjected to matrix logarithmic transformation,

により、ベクトル化する（ｓ８）。ｓ８でベクトル化された、局所領域LocalArea(p)の自己相関行列Σ(p)のベクトルy(p)は、局所領域LocalArea(p)の特徴ベクトル（この発明で言う、局所領域ベクトルに相当する。）である。局所領域LocalArea(p)の特徴ベクトルy(p)は、画素特徴ベクトルがｄ次元のベクトルである場合、（ｄ×（ｄ＋１）／２）次元のベクトルになる。局所領域ベクトル生成機能部３６は、局所領域LocalArea(p)毎に、その局所領域LocalArea(p)の特徴ベクトルy(p)を登録した局所領域ベクトルマップを作成する（ｓ９）。

Thus, vectorization is performed (s8). The vector y (p) of the autocorrelation matrix Σ (p) of the local area LocalArea (p) vectorized in s8 corresponds to the feature vector of the local area LocalArea (p) (referred to as the local area vector in this invention). .) The feature vector y (p) of the local area LocalArea (p) is a (d × (d + 1) / 2) -dimensional vector when the pixel feature vector is a d-dimensional vector. For each local area LocalArea (p), the local area vector generation function unit 36 creates a local area vector map in which the feature vector y (p) of the local area LocalArea (p) is registered (s9).

  Next, the autocorrelation station matrix map in which the global region vector generation function unit 37 registers the autocorrelation matrix M (M (A) to M (G)) of the global region A to G for each of the global regions A to G. Is created (s10). The autocorrelation station matrix map of the global region A to G is

により算出する。［数６］において、NGは、大域領域内に位置する局所領域の数であり、GlobalAreaは大域領域を示す。

Calculated by In [Formula 6], NG is the number of local areas located in the global area, and GlobalArea indicates the global area.

  In s9, in the global region, the autocorrelation matrix M of the feature vector y (p) of the local region LocalArea (p) in the region is calculated. At this time, since the autocorrelation matrix M of the global region is a symmetric matrix, the global region vector generation function unit 37 calculates one of the overlapping elements and uses one calculation result for the other. In this example, the feature vector y (p) of the local area LocalArea (p) has 21 dimensions, so the autocorrelation matrix M of the global area is 21 rows and 21 columns. In addition, there are 210 sets of overlapping elements. That is, in this example, in the calculation of the autocorrelation matrix M in the global region, the calculation for 210 elements can be omitted.

  The global region vector generation function unit 37 creates an autocorrelation matrix map of the global region in which the autocorrelation matrix M (M (A) to M (G)) is associated with each of the global regions A to G.

  Further, the global region vector generation function unit 37 performs matrix logarithmic transformation on the autocorrelation matrix M (M (A) to M (G)) of the global region A to G for each of the global regions A to G ( s11). As described above, since the autocorrelation matrix is a positive definite symmetric matrix, by performing logarithmic transformation of the matrix and vectorizing it, the manifold structure of the autocorrelation matrix is considered in the Euclidean space calculation.

  In s11, matrix logarithmic transformation is performed on the autocorrelation matrix M (M (A) to M (G)) of each of the global regions A to G in the same manner as s7.

Further, the global area vector generation function unit 37 vectorizes the matrix obtained by logarithmic matrix conversion for each of the global areas A to G in consideration of element duplication (s12). When the pixel feature vector is a d-dimensional vector, the feature vectors z (A) to z (G) of the global regions A to G vectorized in s12 are as follows:
It becomes a ((d × (d + 1) / 2) × ((d × (d + 1) / 2) +1) / 2) -dimensional vector. In this example, it is a 231-dimensional vector. The feature vectors z (A) to z (G) of the respective global areas A to G correspond to the global area vector referred to in the present invention.

  The feature quantity generation function unit 38 associates the target person ID for identifying the target person, the feature vectors z (A) to z (G) of each of the global regions A to G, the date and time of the processing target frame image, the imaging location, and the like. The matching data is generated, and the matching data is registered in the feature amount storage unit 6 (s13).

  As described above, the image processing unit 3 calculates the distribution (autocorrelation matrix) of the pixel feature amount in the local region LocalArea (p) as the feature amount of the local region LocalArea (p), and then calculates the local region LocalArea in the global region. The distribution of the feature value of (p) is calculated as the feature value of the global region. The pixel feature distribution in the local area LocalArea (p) of the image is robust against differences in camera performance and imaging environment. The feature quantity of the global area is a feature distribution of the local area LocalArea (p) located in the global area. Therefore, in an object image, even if the distribution amount of pixels with similar feature amounts (pixel feature vectors) such as color and gradient is approximated, the object image has a different pixel distribution in the local region. Different global region vectors are generated.

  Next, the collation process will be described. FIG. 9 is a flowchart showing the matching process.

  The collation unit 4 reads out the collation data for the designated search target person from the feature amount storage unit 6 (s21), and reads out the collation data for the search target person from the feature amount storage unit 6 (s22). The search target person and the search target person verification data are generated by the above-described feature value generation process and stored in the feature value storage unit 6.

  The collation device 1 may be configured such that an operator or the like designates a person to be searched, or the collation unit 4 selects a person to be searched from among persons whose collation data is stored in the feature amount storage unit 6. The configuration may be automatically selected.

  For each of the global regions A to G, the matching unit 4 determines the distance between the search target person's feature vectors z (A) to z (G) and the search target person's feature vectors z (A) to z (G). Is calculated (s23). The collation unit 4 calculates the sum of the distances of the feature vectors z (A) to z (G) calculated for each of the global regions A to G (s24).

  If the sum of the distances calculated in s24 is equal to or less than a predetermined threshold, the matching unit 4 determines that the search target person who has read the matching data in s2 is the search target in which the matching data has been read in s1. Determine (s25, s26). On the contrary, the collation unit 4 calculates the person to be searched for which the collation data is read in s2 and the search object for which the collation data is read in s1, if the sum of distances calculated in s25 is not less than a predetermined threshold. (S25, s27).

  Thus, since this collation apparatus 1 collates an object using the feature-value produced | generated by the feature-value production | generation process mentioned above, it can aim at the improvement of collation precision.

DESCRIPTION OF SYMBOLS 1 ... Collation apparatus 2 ... Image input part 3 ... Image processing part 4 ... Collation part 5 ... Output part 6 ... Feature-value memory | storage part 31 ... Object detection function part 32 ... Object area extraction function part 33 ... Image size normalization function part 34 ... Pixel feature vector map generation function unit
3 5 ... region setting function unit 36 ... local region vector generation unit 37 ... global region vector generation unit 38 ... feature amount generating function unit

Claims (11)

Pixel feature vector extraction that extracts a pixel feature vector for each pixel of an object image obtained by cutting out a region where the object is photographed from a frame image in which the object is captured and converting the extracted region into a predetermined number of pixels And
A first autocorrelation matrix calculating unit that calculates an autocorrelation matrix of a pixel feature vector of a pixel located in the local region for each of a plurality of local regions further set inside the global region set in the object image ;
A local region vector generation unit that generates a local region vector from the autocorrelation matrix of the pixel feature vector calculated by the first autocorrelation matrix calculation unit for each local region;
A second autocorrelation matrix calculating unit that calculates an autocorrelation matrix of a local region vector of the local region located in the global region;
A feature quantity generation unit comprising: a global area vector generation section that generates a global area vector from the autocorrelation matrix of the local area vector calculated by the second autocorrelation matrix calculation section as the global area feature quantity.   The feature according to claim 1, wherein the local region vector generation unit generates a local region vector by performing matrix logarithmic transformation on the autocorrelation matrix of the pixel feature vector calculated by the first autocorrelation matrix calculation unit. Quantity generation unit.   The feature quantity generation unit according to claim 1, wherein the pixel feature vector extraction unit extracts a pixel feature vector including an element related to a luminance value of a pixel.   The feature quantity generation unit according to claim 1, wherein the pixel feature vector extraction unit extracts a pixel feature vector including an element related to an edge of the pixel.   An object image generation unit configured to cut out an area in which the object is imaged from a frame image in which the object is imaged and generate an image obtained by converting the extracted area into a predetermined number of pixels as the object image; Item 5. The feature value generation unit according to any one of Items 1 to 4.   The feature quantity generation unit according to claim 1, wherein the local area set in the global area overlaps with another local area set in the global area.   For each of the plurality of global regions set for the object image, an object feature that uses the global region vector generated by the global region vector generation unit for the global region as a feature amount of the object captured in the object image The feature-value production | generation unit in any one of Claims 1-6 provided with the quantity production | generation part.   The feature amount generation unit according to claim 7, wherein the global area set for the object image overlaps with another global area set for the object image. A feature quantity generation unit according to any one of claims 1 to 8,
Whether the search target object is the search target object using the global region vector generated for the search target object by the feature quantity generation unit and the global region vector generated for the search target object. A collation device having a collation unit for judging. Extract a pixel feature vector of a predetermined dimension from each pixel of the object image obtained by cutting out the region where the object is captured from the frame image where the object is captured and converting the extracted region into a predetermined number of pixels. A pixel feature vector extraction step;
A first autocorrelation matrix calculating step of calculating an autocorrelation matrix of a pixel feature vector of a pixel located in the local region for each of a plurality of local regions further set in the global region set in the object image ;
For each local region, a local region vector generation step for generating a local region vector from the autocorrelation matrix of the pixel feature vector calculated in the first autocorrelation matrix calculation step;
A second autocorrelation matrix calculating step of calculating an autocorrelation matrix of a local region vector of the local region located in the global region;
And a global region vector generation step of generating a global region vector from the autocorrelation matrix of the local region vector calculated in the second autocorrelation matrix calculation step as the global region feature amount. Extract a pixel feature vector of a predetermined dimension from each pixel of the object image obtained by cutting out the region where the object is captured from the frame image where the object is captured and converting the extracted region into a predetermined number of pixels. A pixel feature vector extraction step;
A first autocorrelation matrix calculating step of calculating an autocorrelation matrix of a pixel feature vector of a pixel located in the local region for each of a plurality of local regions further set in the global region set in the object image ;
For each local region, a local region vector generation step for generating a local region vector from the autocorrelation matrix of the pixel feature vector calculated in the first autocorrelation matrix calculation step;
A second autocorrelation matrix calculating step of calculating an autocorrelation matrix of a local region vector of the local region located in the global region;
A feature amount generation program for causing a computer to execute a global region vector generation step for generating a global region vector from the autocorrelation matrix of the local region vector calculated in the second autocorrelation matrix calculation step as the global region feature amount .

Images