JP2010061588A - Feature vector computing device, feature vector computing method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reflect a color with a low color component value appropriately on a feature vector. <P>SOLUTION: A feature vector computing section 18 computes a value of a predetermined increase function on a color component value of at least partial elements out of components of an image, as at least one component of the feature vector of the image, and computes a value of a predetermined decrease function on the color component value of the at least partial elements as at least another component. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a feature vector calculation apparatus, a feature vector calculation method, and a program, and more particularly to calculation of a feature vector of data such as an image.

  A technique is widely used in which a two-dimensional image is composed of a plurality of pixels and one or a plurality of color component values are given to each pixel to describe the contents of the two-dimensional image. Similarly, a method of describing the contents of a three-dimensional image by configuring a three-dimensional image with a plurality of voxels and adding one or a plurality of color component values to each voxel is widely used. Here, the color component value is a value indicating a degree (for example, bright red to dark red) that a color of a component of an image such as a pixel or a voxel has a specific color component.

By the way, the image data itself described by such a method has a large amount of data and is not necessarily suitable for the image matching process. Therefore, various feature vector generation methods have been proposed. For example, higher-order autocorrelation features (see Patent Documents 1 to 3 below) have 1) position invariance that the feature vector does not change even if the position of an object appearing in the image changes, and 2) Since the feature vector of one image and the feature vector of another image are added to the feature vector of the image formed by combining these images, it has an advantageous property that it has an advantageous characteristic. Has been tried.
Japanese Patent No. 2982814 JP 2005-92346 A JP 2006-163452 A

  In general, each component of the image feature vector is given as an increasing function of each color component value of a part or all of the pixels / voxels of the image, including the higher-order autocorrelation features. That is, the larger the value of each color component assigned to the pixel / voxel that is a component of the image, the larger the value of each component of the feature vector, and thus the larger the norm (size) of the feature vector. Conversely, the smaller each color component value, the smaller the value of each component of the feature vector, and hence the smaller the norm of the feature vector. Therefore, a pixel / voxel having a small color component value has less influence on the image feature vector than a pixel / voxel having a large color component value, and according to the conventional method, a color having a small color component value is not appropriately reflected in the feature vector. There was a problem.

  The present invention has been made in view of the above problems, and an object thereof is to provide a feature vector calculation device, a feature vector calculation method, and a program capable of appropriately reflecting a color having a small color component value in a feature vector. There is. Another object of the present invention is to provide a feature vector calculation device, a feature vector calculation method, and a program capable of appropriately exerting an influence of an element having a small value on a feature vector of data composed of a plurality of elements. It is in.

  In order to solve the above-described problem, the feature vector calculation apparatus according to the present invention indicates the degree to which each of a plurality of elements constituting an image has one or more color components. A color component value acquisition unit that acquires a color component value; and a feature vector calculation unit that calculates a feature vector of the image. The feature vector calculation unit includes the plurality of feature vectors as at least one component of the feature vector. Principal component calculating means for calculating a value of a predetermined increase function relating to color component values of at least some of the elements, and color component values of the at least some of the elements as at least one other component of the feature vector And auxiliary component calculating means for calculating a value of a predetermined decreasing function for.

  In addition, the feature vector calculation method according to the present invention obtains a color component value indicating the degree to which the color of the element has one or more color components for each of the plurality of elements constituting the image. A color component value acquisition step; and a feature vector calculation step of calculating a feature vector of the image. The feature vector calculation step includes at least a part of the plurality of elements as at least one component of the feature vector. A principal component calculating step for calculating a value of a predetermined increase function relating to the color component value of the element, and a predetermined decrease function relating to the color component value of the at least some element as at least one other component of the feature vector. And an auxiliary component calculating step for calculating a value.

  Also, the program according to the present invention obtains a color component value for obtaining a color component value indicating the degree to which the color of the element has one or a plurality of color components for each of the plurality of elements constituting the image. Acquisition means; principal component calculation means for calculating a value of a predetermined increase function relating to color component values of at least some of the plurality of elements as at least one component of the feature vector of the image; and This is a program for causing a computer to function as auxiliary component calculating means for calculating a value of a predetermined decreasing function relating to the color component values of at least some of the elements as at least one other component. The program may be stored in a computer-readable storage medium such as a CD-ROM or DVD-ROM.

  According to the present invention, a predetermined increase function value related to a color component value is used as at least one component (principal component) of a feature vector, and a predetermined value related to the same color component value is used as at least one other component (auxiliary component). The value of the decreasing function is used. For this reason, even if the color component value is small and therefore the main component of the feature vector is small, the auxiliary component is large instead. Thus, according to the present invention, a color having a small color component value can be appropriately reflected in the feature vector.

  In one aspect of the invention, the predetermined decreasing function has a value range equal to the predetermined increasing function. In this way, the influence on the feature vector can be made equal between the normal component and the auxiliary component.

  In the aspect of the invention, the auxiliary component calculating unit calculates an auxiliary color component value that is a value of a decrease function for a predetermined auxiliary color component value related to the color component values of the at least some elements. Then, the auxiliary color component value is substituted into the predetermined increase function to calculate the at least one other component of the feature vector. At this time, the reduction function for the predetermined auxiliary color component value may be determined so that a range that the auxiliary color component value can take and a range that the color component value can take. Further, the color component value may take a value between zero and a predetermined value, and the decreasing function for the predetermined auxiliary color component value may subtract the color component value from the predetermined value.

  Further, the auxiliary component calculation means may give zero as each color component value of the element and the auxiliary color component value when an object is not represented in the element of the image. By doing this, it is possible to generate different feature vectors when the object is not represented in the image element and when the object element is given a color with all color component values being zero, Both cases can be distinguished by feature vectors.

  In one aspect of the present invention, the feature vector calculation unit calculates a correlation value between a part of the plurality of elements and a predetermined element pattern while shifting the part of the range, The at least one component and the at least one other component of the feature vector are calculated by integrating the calculated correlation values. In this way, position invariance and additivity of feature vectors can be realized as in the case of higher-order local autocorrelation features.

  Further, according to an aspect of the present invention, distance calculation means for calculating a distance between a given vector and the feature vector is further included. In this way, the similarity between the feature vector calculated by the present invention and a given vector can be obtained.

  The feature vector calculation apparatus according to the present invention includes a value acquisition unit that acquires each value of a plurality of elements constituting data, and a feature vector calculation unit that calculates a feature vector of the data, and the feature Vector calculation means includes, as at least one component of the feature vector, principal component calculation means for calculating a value of a predetermined increase function related to the values of at least some of the plurality of elements; And auxiliary component calculating means for calculating a value of a predetermined decreasing function related to the values of at least some of the elements as at least one component.

  Further, the feature vector calculation method according to the present invention includes a value acquisition step of acquiring each value of a plurality of elements constituting the data, and a feature vector calculation step of calculating a feature vector of the data, The vector calculation step includes, as at least one component of the feature vector, a principal component calculation step of calculating a value of a predetermined increase function related to the values of at least some of the plurality of elements; And an auxiliary component calculating step of calculating a value of a predetermined decreasing function related to the values of at least some of the elements as at least one component.

  Further, the program according to the present invention includes a value acquisition means for acquiring values of a plurality of elements constituting data, and at least a part of the plurality of elements as at least one component of the feature vector of the data A principal component calculating means for calculating a value of a predetermined increase function relating to the value of the auxiliary component, and an auxiliary component for calculating a value of a predetermined decreasing function relating to the value of at least some of the elements as at least one other component of the feature vector A program for causing a computer to function as calculation means.

  According to the present invention, as at least one component (principal component) of a feature vector, a value of a predetermined increase function relating to the value of a component of data is used, and as at least one other component (auxiliary component), it relates to the same value. A value of a predetermined decreasing function is used. For this reason, even if the value of a certain element is small and therefore the normal component of the feature vector is small, the auxiliary component is large instead. Thus, according to the present invention, the influence of an element having a small value can be appropriately exerted on the feature vector of data composed of a plurality of elements.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a functional block diagram of a feature vector calculation apparatus according to an embodiment of the present invention. A feature vector calculation apparatus 10 shown in the figure includes an initial image data storage unit 12, a voxel value extension unit 14, an extended image data storage unit 16, a feature vector calculation unit 18, a feature vector storage unit 20, and a distance calculation unit 22. It is out. These functional blocks are realized by executing programs on various computers, for example. Alternatively, each functional block may be implemented by hardware. The program may be read into the computer by a computer-readable medium such as a CD-ROM or DVD-ROM, or may be downloaded to the computer via a communication network such as the Internet.

The initial image data storage unit 12 stores image data for which a feature vector is calculated, and includes, for example, a hard disk storage device or a memory. Here, the three-dimensional image data shown in FIG. 2 is a feature vector calculation target. The three-dimensional image 34 shown in the figure is composed of M1 × M2 × M3 voxels 34a. The voxel 34a is a small cubic image element. A flower pot 32 is shown at the center of the three-dimensional image 34, and a space 30 is shown around it. The content of the three-dimensional image 34 may be anything. For example, the three-dimensional image 34 may be a three-dimensional image in which images of frames constituting a moving image are arranged in the time direction. In this way, the feature of the moving image can be extracted. Each voxel is provided with color information, and the color information includes color component values R, G, and B indicating the degree to which the color represented by the voxel has red, green, and blue color components. Yes. Here, a three-dimensional color vector f (r) indicating the color of the voxel at the position r is expressed as the following equation (1). In Equation (1), r is a three-dimensional vector that specifies the position of the voxel, and I _R (r) is a value (red component value) indicating the degree that the color of the voxel at the position r has a red component. , I _G (r) is a value (green component value) indicating the degree that the color of the voxel at the position r has a green component, and I _B (r) indicates the degree that the color of the voxel at the position r has a blue component. Value (blue component value). The values of I _R (r), I _G (r), and I _B (r) are all 0 or more and 1 or less. The initial image data storage unit 12 stores color vectors f (r) for all positions r in the image 34. In this case, three components of red, green, and blue are used to express the color represented by each voxel. However, the present invention is not limited to this, and is used for HSV and television signals used in computer graphics. The color represented in each voxel may be expressed by other color components such as YIQ. The initial image data storage unit 12 may store a flag indicating whether or not an object is represented in each voxel.

  The voxel value extending unit 14 extends the color vector f (r) stored in the initial image data storage unit 12 to a 6-dimensional extended color vector F (r) as shown in the following equation (2), Are stored in the extended image data storage unit 16 including a hard disk storage device and a memory.

The first, third and fifth components of the extended color vector F (r) are the same as the components contained in the original color vector f (r), I _R (r), I _G (r) and I _B (r). On the other hand, the second, fourth and sixth components are obtained by subtracting each value from 1 which is the maximum value of I _R (r), I _G (r) and I _B (r), They become auxiliary color component values for the original I _R (r), I _G (r), and I _B (r). The auxiliary color component value also takes a value between 0 and 1. Specifically, the voxel value expansion unit 14 reads the values of the components of the color vector f (r) stored in the initial image data storage unit 12, calculates the value obtained by subtracting them from 1, and expands each of them. The second, fourth, and sixth components of the color vector F (r) are stored in the extended image data storage unit 16. Each component of the read color vector f (r) is also stored in the extended image data storage unit 16 as the first, third, and fifth components of the extended color vector F (r). Note that the second component of the extended color vector F (r) indicates the degree to which the color of the voxel at the position r does not have a red component, and the fourth component shows the degree to which the color component does not have a green component. The sixth component indicates a degree not having a blue component. Here, although the value obtained by subtracting the components of 1 from the color vector f (r) and the auxiliary color component ^{_{value, (1-I R (r}} ) 2) 1/2, (1-I G (r) 2 ) ^1/2 , (1-I _B (r) ² ) ^1/2, etc. Other color component values may be used as auxiliary color component values. Also in this case, it is desirable that the value range of the auxiliary color component value matches the value range of the original color component value.

  The voxel value extending unit 14 performs all color component values and auxiliary colors for voxels in which no object is represented, such as voxels located inside the flower pot 32 in the three-dimensional image 34 or voxels located in the space 30. Set the component value to zero. When a flag indicating whether or not an object is represented in each voxel is stored in the initial image data storage unit 12, whether or not an object is represented in each voxel is referred to by referring to these flags. It can be easily judged. Thus, by setting all component values of the extended color vector F (r) to zero for voxels in which no object is represented, it is possible to prevent voxels in which no object is represented from affecting the feature vector.

  The feature vector calculation unit 18 calculates a higher-order local autocorrelation feature extended from the extended color vector F (r) stored in the extended image data storage unit 16 so that the features of a three-dimensional and colored image can be described. Is. Specifically, the correlation value between each image pattern of a small region composed of 3 × 3 × 3 voxels illustrated in FIGS. 3 and 4 and a local region of the three-dimensional image 34 is calculated, Correlation values are sequentially calculated for all local regions while shifting the local regions within the three-dimensional image 34. Then, by integrating these correlation values, each component of the feature vector is obtained. For example, FIG. 3A shows an image pattern in which a voxel having a red component value of 1 is arranged at the center (target pixel), and voxels of an arbitrary color are arranged around the center, In FIG. 6B, a voxel having a red auxiliary component value of 1 (= a red component value of 0) is arranged at the target pixel, and an image pattern in which voxel images of arbitrary colors are arranged around the pixel. It is shown. (C) and (d) are image patterns created for the green component in the same manner, and (e) and (f) are image patterns created for the blue component in the same manner. These image patterns in which the color of only the pixel of interest is designated are referred to as 0th-order patterns. There are six zeroth-order patterns, the same as the number of dimensions of the extended color vector F (r).

  FIG. 4 shows a part of the primary pattern. In FIGS. 4A to 4F, a voxel having a red component value of 1 is arranged at the target pixel, and a red component value, a red auxiliary component value, a green component value, A voxel having an auxiliary component value, a blue component value, and a blue auxiliary component value of 1 is arranged, and an image pattern in which voxels of an arbitrary color are arranged around is shown. In addition to these pixels of interest, an image pattern in which one of the surrounding colors is designated is called a primary pattern. There are 468 (= 6 × 6 × 13) primary patterns depending on the number of dimensions of the extended color vector F (r) (= 6) and the variation in the position of the voxel for which the color is designated (= 13). .

  The integrated value of the local correlation between each Nth-order pattern and the three-dimensional image 34, that is, the Nth-order higher-order local autocorrelation feature is expressed by the following equation (3).

Here, a _i (i = 1 to N) is a three-dimensional vector indicating the position of the voxel for which the color is designated in the Nth-order pattern other than the pixel of interest. Moreover, x surrounded by a circle indicates a tensor product. In the computer implementation, the integration is replaced with the sum of the positions r of some or all of the voxels in the three-dimensional image 34. For example, in the three-dimensional image 34 shown in FIG. 2, the sum of the positions r of (M1-2) × (M2-2) × (M3-2) voxels may be calculated. In the above equation (3), the feature is indicated by a tensor, but the element is used as a component of the feature vector of the image.

  In the present embodiment, only for the 0th-order pattern and the primary pattern, the integrated value of the local correlation with the three-dimensional image 34 is obtained and used as the component of the feature vector. When the integrated value of the local correlation between the zeroth-order pattern and the three-dimensional image 34 is specifically written, the following equation (4) is obtained. The left side x0 of Expression (4) is a 6-dimensional vector, and each component thereof is used as the first to sixth components of the feature vector.

When the integrated value of the local correlation between the primary pattern and the three-dimensional image 34 is specifically written, the following equation (5) is obtained. Here, a _i (i = 1 to 13) is a three-dimensional vector indicating the position of the voxel for which the color is designated in addition to the target pixel in each primary pattern. The left side x ₁ (a _i ) of Expression (5) is a 6 × 6 matrix, and when the left side x ₁ (a _i ) is calculated for all i = 1 to 13, a total of 468 matrix elements are obtained. . These are used as the seventh to 474th components of the feature vector of the three-dimensional image 34.

In the present embodiment, the correlation between different color components in the color of the pixel of interest is further extracted as a feature. X ₁ ′ shown in the following equation (6) indicates such a feature. The left side x ₁ ′ is a 6 × 6 matrix, except for 15 matrix elements having the same value, and the remaining 21 matrix elements are further replaced with the 475th to 495th components of the feature vector of the three-dimensional image 34. I will use it as

  The feature vector calculated by the feature vector calculation unit 18 is stored in a feature vector storage unit 20 configured by a hard disk storage device or a memory. Then, when a feature vector of a three-dimensional image to be compared is input to the distance calculation unit 22 from the outside, the distance between the input feature vector and the feature vector stored in the feature vector storage unit 20 is calculated. The distance may be various types of distances such as Euclidean distance and cosine distance. If the calculated distance is larger than the predetermined threshold, the determination result is output as dissimilar, and if the calculated distance is equal to or smaller than the predetermined threshold, the determination result is output as similar.

  When the feature vector of the three-dimensional image 34 calculated in the present embodiment is specifically written, the following equation (7) is obtained.

As can be seen from equation (7), ΣI _R (r), which is one component (principal component) of the feature vector, is given by an increasing function for the red component value I _R (r). Further, Σ (1-I _R (r)) which is another component (auxiliary component) is given by a decreasing function with respect to the red component value I _R (r). Similarly, ΣI _R (r) I _R (r + a ₁ ), which is another component (principal component) of the feature vector, is given by an increasing function for the red component value I _R (r). Further, Σ ((1-I _R (r) I _R (r + a ₁ )), which is another component (auxiliary component), is given by a decreasing function with respect to the red component value I _R (r). When one component (principal component) of a feature vector is calculated with an increasing function of a certain color component value, another component (auxiliary component) is always calculated with a decreasing function. In addition, a value (auxiliary component) obtained by substituting the auxiliary color component value of the color together with a value (main component) obtained by substituting the color component value of a certain color is used as a component of the feature vector. Therefore, the range of possible values of these two components is equal.

According to this embodiment, in addition to the value of the increasing function of I _R (r), I _G (r), or I _B (r), the value of the decreasing function is also a component of the feature vector. _{Therefore, I R (r), I} G (r) or I _{B (r)} is small, therefore even if certain components of the feature vector becomes small, is calculated by the other ingredients, namely decreasing function instead The component value increases. For this reason, the norm of the feature vector is not excessively small, and accordingly, a color having a small color component value can be appropriately reflected in the feature vector.

  The present invention is not limited to the above embodiment. For example, in the above-described embodiment, an example in which the present invention is applied to higher-order local autocorrelation features has been described. However, the present invention is not limited thereto, and the present invention can be applied to other types of features in the same manner. In addition, the present invention can be similarly applied not only to a three-dimensional image but also to a two-dimensional image, and also to a monochrome image and a grayscale image as well as a color image. Furthermore, although the method for calculating the feature vector of an image has been described as an example here, the feature vector of data composed of a plurality of elements, for example, data in which each element takes a value not less than 0 and not more than a predetermined value, is the same as described above. Of course, it can be calculated.

It is a functional block diagram of the feature vector calculation device concerning the embodiment of the present invention. It is a figure which shows an example of the three-dimensional image used as the calculation object of a feature vector. It is a figure which shows a 0th-order pattern. It is a figure which shows a part of primary pattern.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 10 Feature vector calculation apparatus, 12 Initial image data storage part, 14 Voxel value expansion part, 16 Extended image data storage part, 18 Feature vector calculation part, 20 Feature vector storage part, 22 Distance calculation part, 30 Space, 32 Flower pot, 34 3D image, 34a voxel.

Claims (13)

Color component value acquisition means for acquiring a color component value indicating the degree to which the color of the element has one or more color components for each of a plurality of elements constituting the image;
Feature vector calculation means for calculating a feature vector of the image,
The feature vector calculation means includes:
Principal component calculating means for calculating a value of a predetermined increase function related to color component values of at least some of the plurality of elements as at least one component of the feature vector;
An auxiliary component calculating means for calculating a value of a predetermined decreasing function related to color component values of the at least some elements as at least one other component of the feature vector;
A feature vector calculation apparatus characterized by the above. In the feature vector calculation device according to claim 1,
The predetermined decreasing function has a range equal to the predetermined increasing function;
A feature vector calculation apparatus characterized by the above. In the feature vector calculation device according to claim 1 or 2,
The auxiliary component calculation means calculates an auxiliary color component value that is a value of a decrease function for a predetermined auxiliary color component value related to the color component values of the at least some elements, and calculates the auxiliary color component value as the predetermined color value. Substituting in an increasing function of the at least one other component of the feature vector;
A feature vector calculation apparatus characterized by the above. In the feature vector calculation device according to claim 3,
The decreasing function for the predetermined auxiliary color component value is determined such that a range that the auxiliary color component value can take and a range that the color component value can take.
A feature vector calculation apparatus characterized by the above. In the feature vector calculation device according to claim 4,
The color component value takes a value between zero and a predetermined value,
The decreasing function for the predetermined auxiliary color component value subtracts the color component value from the predetermined value;
A feature vector calculation apparatus characterized by the above. In the feature vector calculation device according to any one of claims 3 to 5,
The auxiliary component calculation means, when an object is not represented in the element of the image, to give zero as each color component value of the element and their auxiliary color component value,
A feature vector calculation apparatus characterized by the above. In the feature vector calculation device according to any one of claims 1 to 6,
The feature vector calculation means calculates a correlation value between a part of the plurality of elements and a predetermined element pattern while shifting the part of the range, and integrates the calculated correlation values. Calculating the at least one component and the at least one other component of the feature vector,
A feature vector calculation apparatus characterized by the above. In the feature vector calculation device according to any one of claims 1 to 7,
A distance calculating means for calculating a distance between a given vector and the feature vector;
A feature vector calculation apparatus characterized by the above. A color component value acquisition step for acquiring a color component value indicating the degree to which the color of the element has one or more color components for each of the plurality of elements constituting the image;
A feature vector calculation step of calculating a feature vector of the image,
The feature vector calculation step includes:
A principal component calculating step of calculating a value of a predetermined increase function related to color component values of at least some of the plurality of elements as at least one component of the feature vector;
An auxiliary component calculating step of calculating a value of a predetermined decreasing function related to color component values of the at least some elements as at least one other component of the feature vector,
A feature vector calculation method characterized by the above. Color component value acquisition means for acquiring a color component value indicating the degree to which the color of the element has one or more color components for each of a plurality of elements constituting the image;
As at least one component of the feature vector of the image, principal component calculating means for calculating a value of a predetermined increase function relating to color component values of at least some of the plurality of elements, and at least another of the feature vectors A program for causing a computer to function as auxiliary component calculation means for calculating a value of a predetermined decreasing function relating to the color component values of at least some of the elements as one component. Value acquisition means for acquiring values of each of a plurality of elements constituting the data;
Feature vector calculation means for calculating a feature vector of the data,
The feature vector calculation means includes:
Principal component calculating means for calculating a value of a predetermined increase function related to the values of at least some of the plurality of elements as at least one component of the feature vector;
Auxiliary component calculating means for calculating a value of a predetermined decreasing function related to the values of at least some of the elements as at least one other component of the feature vector;
A feature vector calculation apparatus characterized by the above. A value acquisition step of acquiring each value of a plurality of elements constituting the data;
A feature vector calculation step of calculating a feature vector of the data,
The feature vector calculation step includes:
A principal component calculating step of calculating a value of a predetermined increase function related to values of at least some of the plurality of elements as at least one component of the feature vector;
An auxiliary component calculating step of calculating a value of a predetermined decreasing function related to the values of at least some of the elements as at least one other component of the feature vector;
A feature vector calculation method characterized by the above. Value acquisition means for acquiring each value of a plurality of elements constituting the data;
Principal component calculation means for calculating a value of a predetermined increase function related to the values of at least some of the plurality of elements as at least one component of the feature vector of the data; and at least one other of the feature vector A program for causing a computer to function as auxiliary component calculation means for calculating a value of a predetermined decrease function relating to the values of at least some of the elements as components.

Images