JP6547280B2 - Setting device, information classification device, classification plane setting method of setting device, information classification method of information classification device and program - Google Patents

Description

  The present invention relates to a setting apparatus, an information classification apparatus, a classification plane setting method for the setting apparatus, an information classification method for an information classification apparatus, and a program for setting a classification plane used for information classification in an information classification apparatus.

  When classifying similar or dissimilar such as fingerprints and faces, an information classification device using a support vector machine (SVM) is used. The SVM constructs in advance a classification surface (a function representing a classification condition for classifying information) for classifying information using known learning information as to how unknown information should be classified. , Classify unknown information using its classification surface. SVMs can be roughly divided into linear SVMs and non-linear SVMs, but non-linear SVMs that classify information using a technology that maps information to a feature space can classify more complex information more accurately.

  Non-Patent Document 1 discloses a functional expression in which each element of a feature vector is raised to a power as a kernel used to construct a classification surface to be set to a non-linear SVM. The kernel is a functional expression for obtaining the relationship between feature vectors extracted from information of a classification target. Non-Patent Document 1 discloses a technique for increasing the classification accuracy of an image by using the disclosed kernel.

Sabri Boughorbel, 2 others, "GENERALIZED HISTOGRAM INTERSECTION KERNEL FOR IMAGE RECOGNITION", International Conference on Image Processing (ICIP), September 2005

However, the kernel disclosed in Non-Patent Document 1 performs calculation processing for raising the value of each element of the feature vector. The processing of the exponentiation calculation is heavy in calculation processing. Therefore, when the technology disclosed in Non-Patent Document 1 is used, there is a problem that when the number of elements of the feature vector is increased, the number of times of the exponentiation process in which the processing load is large is also increased and the processing load is increased.
The present invention has been made in view of the above problems, and a setting device, an information classification device, and a setting device classification in which the processing load of the power calculation of the kernel is small even if the number of elements of the feature vector acquired from information increases An object of the present invention is to provide a plane setting method, an information classification method of an information classification device, and a program.

In order to achieve the above object, the setting device according to the present invention indicates a plurality of learning information whose classification is known, a plurality of test information whose classification is known, and an allowance for misclassification of the learning information. Input means for inputting a plurality of regularization parameters and a plurality of exponentiation parameters for raising the similarity between feature vectors, and an acquisition means for acquiring feature vectors respectively corresponding to the learning information and the test information And learning means for constructing a plurality of classification planes for classifying the learning information for each combination of the regularization parameter and the exponentiation parameter based on the power of the similarity between the feature vectors, and the learning means Selecting one of the plurality of classification planes constructed in step (d) and classifying the feature vector corresponding to the test information generated by the acquisition means, and dividing the feature vector corresponding to the test information The classification rate which represents the proportion classified into the known classification state to be determined is determined, and further, the classification surface which is selected successively is updated several times, and the classification surface determined the highest by the classification ratio determined by the test means. And a setting unit configured to set the classification plane selected by the selection unit in the information classification apparatus.
Further, in order to achieve the above object, an information classification device according to the present invention is an information classification means for classifying unknown information different from learning information and test information according to the classification surface constructed by the setting device. And the like.

  According to the present invention, even if the number of elements of a feature vector increases, the processing load of kernel power calculation can be reduced.

It is a block diagram of the information classification device concerning Embodiment 1 of the present invention. It is a function block diagram of the information classification device shown in FIG. It is a figure for demonstrating the process which extracts a feature vector from an image, and normalizes the extracted feature vector, (a) shows the example which divided | segmented the image into the block (5 vertical bars, 5 horizontal bars), (B) is a figure for demonstrating normalization of a feature vector. It is a conceptual diagram for demonstrating classifying a feature vector into a binary, (a) is a figure explaining the case where a classification surface can not be set easily, when (b) can easily set a classification surface. is there. It is a figure for demonstrating the example of calculation of a kernel, (a) is a feature vector of a training image for normalization, (b) is a feature vector of another training image for normalization, (c) is a kernel (D) shows the bar graph of (a), (e) shows the bar graph of (b), and (f) shows the overlapping part of (d) and (e). It is a figure for demonstrating the structure of a classification | category surface, (a) is a case where it can not linear isolate | separate, (b) shows the case where it maps to the linearly separable feature space. It is a flowchart figure for demonstrating the setting process of the classification | category surface which the information classification apparatus which concerns on Embodiment 1 of this invention performs. It is a figure for demonstrating the processing load reduction effect of the power calculation with respect to a prior art. It is a functional block diagram of the setting apparatus which sets a classification | category surface to the information classification apparatus which concerns on Embodiment 2 of this invention. It is a function block diagram of the information classification device which concerns on Embodiment 2 of this invention.

  Hereinafter, an information classification device, a setting method of the information classification device, and a program according to an embodiment of the present invention will be described with reference to the drawings. In this embodiment, the case where the information classification device is applied to classification of image information will be described.

(Embodiment 1)
The hardware configuration of the information classification device according to the first embodiment of the present invention will be described with reference to FIG. The information classification device 11 according to the first embodiment includes an input unit 12, a display unit 13, an input / output interface (input / output I / F) unit 14, a storage unit 15, a read only memory (ROM) 16, and a random access memory (RAM). And 17. a control unit 18.
The input unit 12 is configured by a keyboard or the like to input information such as an instruction for specifying a learning image, a test image, and a classification target image by a user, and a parameter necessary for information classification processing.
The display unit 13 displays various information such as input parameter values, identification rates to be described later, and classified images, and is configured by a liquid crystal display or the like.
The input / output I / F unit 14 has an interface function with an external memory, a personal computer, and the like, and is configured of a USB (Universal Serial Bus) interface, a compact disk drive, and the like.
The storage unit 15 stores a learning image, a test image, a classification target image, feature vectors of images analyzed in the past and images to be described later, classification plane information used for classification of feature vectors, etc. Be done.
The ROM 16 stores a non-linear SVM processing program executed by the control unit 18 including a Bof (Bag-of-Features) program for extracting feature vectors from an image.
The RAM 17 is used as a work area of the control unit 18.

  The control unit 18 includes a processor or the like, reads out the nonlinear SVM processing program stored in the ROM 16 to the RAM 17, and functions as an information classification device by executing the program. As shown in FIG. 2, the control unit 18 functionally includes an acquisition unit 21, a normalization unit 22, a learning unit 23, a test unit 24, a selection unit 25, a setting unit 26, and a classification unit 27.

  The acquisition unit 21 reads from the ROM 16 a Bof program for extracting feature vectors of an image, and extracts feature vectors from a learning image group, a test image group, and a classification target image group to be analyzed. The acquisition unit 21 stores the extracted feature vector in the RAM 17. When the amount of information is large, it may be stored in the storage unit 15.

A feature vector is a quantification of features extracted from an image. In this embodiment, an image is divided into blocks, pixels in the block are determined, and the frequency of colors forming the block is extracted as a feature vector. That is, it counts which color has which block, and counts the count value as an element of the feature vector.
In order to facilitate understanding, the five colors (pink, red, blue, white, black) are divided into blocks of 5 squares vertically and 5 squares horizontally as shown in FIG. The frequency of whether it exists in is extracted as an element of the feature vector. In the image of the face illustrated in FIG. 3 (a), (pink) exists in 10 blocks in 25 blocks, (red) exists in 2 blocks, and (blue) is 1 In the block, (white) is present in 7 blocks and (black) is present in 5 blocks. What extracted this color frequency as an element of a feature vector is shown in the 2nd line of FIG.3 (b).

Returning to FIG. 2, the normalization unit 22 normalizes the feature vector acquired by the acquisition unit 21 with the L2 norm. The L2 norm is the square root of the sum of squares of each element of the feature vector. In this normalization, for example, when classifying the images, if the sizes of the images to be compared are different but they are similar, they should be classified as similar as the images, but the sizes of the comparison target images are different. It has a function to prevent it from being classified as dissimilar.
The sum of squares of each element of the feature vector described in the second line of FIG. 3B is (179), and the square root thereof is (13. 38). Then, on the third line of FIG. 3B, each element of the feature vector obtained by normalizing each element of the feature vector described on the second line with the square root of the sum of squares of each element of the feature vector is described ing.

  Returning to FIG. 2, the learning unit 23 constructs a classification surface for classifying the normalized feature vectors of the learning image, using the on-line nonlinear SVM. The learning unit 23 constructs this classification surface for each combination of a regularization parameter C indicating the tolerance of classification error described later and the power parameter α of the kernel. For example, if there are 100 combinations of the regularization parameter C and the exponentiation parameter α, 100 classification planes are constructed.

The test unit 24 classifies a feature vector corresponding to a known test image using the classification surface constructed by the learning unit 23. Then, the test unit 24 obtains an identification rate indicating a rate at which the feature vector is classified into a known group (class). The test unit 24 obtains the identification rate for each of the plurality of classification planes constructed by the learning unit 23.
For example, if there are 100 test images for Mr. A, how many of the 100 test images are correctly classified as Mr. A to find out the identification rate. If 99 out of 100 sheets are classified as Mr. A, the discrimination rate is 99%, and if 95 sheets are classified as Mr. A, the discrimination rate is 95%. If the number of classification planes constructed by the learning unit 23 is 100, the test unit 24 obtains an identification rate for each of the 100 classification planes.

The selection unit 25 selects the classification plane having the highest discrimination rate among the discrimination rates obtained for each classification plane in the test unit 24.
The setting unit 26 sets the classification surface having the highest discrimination rate selected by the selection unit 25 as the nonlinear SVM.

  The classification unit 27 classifies the feature vector corresponding to the image of the classification target whose method of classification is unknown using the non-linear SVM for which the classification plane is set by the setting unit 26. Specifically, the user instructs to supply the classification target image group to the acquisition unit 21, and the acquisition unit 21 extracts the feature vector of each classification target image. The normalization unit 22 normalizes the extracted feature vector and supplies the feature vector to the classification unit 27. The classification unit 27 classifies the classification target image by classifying the feature vector corresponding to the input classification target image using the classification plane set in the nonlinear SVM.

  The configuration of the information classification device 11 has been described above. Next, detailed functions of the learning unit 23 will be described. In the present embodiment, an example will be described in which feature vectors of Mr. A's image are represented by white circles and classified into group A, and feature vectors of other images are represented by black circles and classified into group B. In the example shown in FIG. 4A, the feature vectors represented by white circles are distributed in a clump and the feature vectors represented by black circles are also distributed in a clump. Can be easily built. However, in the example shown in FIG. 4B, since the feature vector displayed as white circles and the feature vector displayed as black circles are distributed in an intricate manner, it is impossible to easily construct a classification plane.

  As described above, when the classification surface can not be easily constructed, the learning unit 23 has a function of allowing classification errors to a certain extent and constructing the classification surface. In this embodiment, the degree to which this classification error is permitted is represented by the regularization parameter C. When the regularization parameter C is increased, a classification surface that does not tolerate classification errors is constructed, and when the regularization parameter C is decreased, classification errors are allowed to some extent, and a classification surface is constructed.

  The classification surface indicated by a solid line in FIG. 4B is a case where the regularization parameter C is large and the classification error is not permitted, and the classification surface has a complicated shape. Because the classification plane is intricate, the distance between the feature vector and the classification plane (classification margin) tends to be smaller. On the other hand, the classification plane indicated by the dotted line is a case where the regularization parameter C is small and the classification error is quite tolerated. Classification planes can be constructed with simple curves, and the classification margin between feature vectors that are not misclassified and classification planes tends to be large.

この分類面を表す識別関数を式（２）で表す。

The learning unit 23 constructs a classification surface using the learning function of the nonlinear SVM. In the nonlinear SVM, the learning feature vector x _i which is a linear non-separable distribution as shown in FIG. 6 (a) can be linearly separated φ (x _i ) as shown in FIG. 6 (b) using non-linear space transformation. Transform to the feature space of to construct a classification surface.

A discrimination function representing this classification surface is expressed by equation (2).

ここで、学習用画像の特徴ベクトルｘ_ｉに関する教師信号ｙ_ｉを次のように定義する。

Ａ氏の画像を分類する場合では、ｘ_ｉがＡ氏の画像である場合はｙ_ｉ＝１であり、ｘ_ｉがＡ氏の画像ではない場合はｙ_ｉ＝−１である。 The feature vector x _i corresponding to the learning image is i = 1 to n. This feature vector x _i is classified into a group (class) A and a group (class) B. If a weight vector m and a bias term b are obtained for this feature vector so that g (x) satisfies the following condition, a discrimination function representing a classification surface can be constructed. With respect to this feature vector, the weight vector m and the bias term b are adjusted so that g (x) satisfies the following condition.

Here, the teacher signal y _i regarding the feature vector x _i of the learning image is defined as follows.

In the case of classifying the image of Mr. A, if x _i is the image of Mr. A is y _{i =} 1, if x _i is not the image of Mr. A is y _{i =} -1.

学習部２３は、図６（ｂ）に実線で示す分類面（識別関数）から最寄りの点（φ（ｘ_Ａ１），φ（ｘ_Ｂ１））までの距離（分類マージンＬ）を最大化するように重みベクトルｍとバイアス項ｂを求めて分類面を構築する。図６（ｂ）に示す分類マージンＬは、この最寄りの点φ（ｘ_Ａ１）を含み分類面と平行する関数Ｈ１と、点φ（ｘ_Ｂ２）を含み分類面と平行する関数Ｈ２とから求めることができる。

If Expression (3) is rewritten using Expression (4), Expression (5) is obtained.

The learning unit 23 maximizes the distance (classification margin L) from the classification surface (classification function) indicated by the solid line in FIG. 6B to the nearest points (φ (x _A1 ), φ (x _B1 )). The weight vector m and the bias term b are obtained to construct a classification surface. The classification margin L shown in FIG. 6B is obtained from the function H1 including the nearest point φ (x _A1 ) parallel to the classification plane and the function H2 including the point φ (x _B2 ) parallel to the classification plane be able to.

この分類マージンＬが最大化するために、||ｍ||を最小化すれば良い。この式（７）を等価である式（８）に置き換える。

Since the classification margin L is half of the distance between H1 and H2, when the two are subtracted to obtain the margin L, y _i m (φ (x _A1 ) −φ (x _B1 )) = 2. If both sides are divided by the weight vector size || m ||, then y _i (φ (x _A1 ) −φ (x _B1 )) = 2 / || m || Since it exists, it becomes Formula (7).

In order to maximize the classification margin L, it is sufficient to minimize || m ||. This equation (7) is replaced with the equivalent equation (8).

λは、ラグランジュの未定乗数である。これをｍ及びｂで偏微分して０とおくと、

式（１０）より、

ラグランジュ関数ＬｐをＦ（λ）と書き換えて、式（１１）及び式（１２）を代入すると、式（１３）となる。

If the problem of maximizing equation (8) is rewritten using Lagrange's undetermined multiplier method, equation (9) is obtained.

λ is Lagrange's undetermined multiplier. If this is differentially differentiated by m and b and set to 0,

From equation (10),

If the Lagrange function Lp is rewritten as F (λ) and the equations (11) and (12) are substituted, the equation (13) is obtained.

ここで、内積部分φ（ｘ_ｉ）φ（ｘ_ｊ）をカーネルｋ（ｘ_ｉ，ｘ_ｊ）で置き換えると、式（１５）となる。

Substituting equation (12) into equation (13) results in equation (14).

Here, when the inner product part φ (x _i ) φ (x _j ) is replaced by the kernel k (x _i , x _j ), equation (15) is obtained.

Under the constraint of the regularization parameter C ≧ λ _i 00, the weight vector λ at which this F (λ) becomes maximum is determined, thereby determining m in equation (12).

式（１６）に、図６（ｂ）に示す分類マージンを決める特徴ベクトルｘ_Ａ１，ｘ_Ｂ１を代入することにより、バイアス項ｂが決まる。 If equation (12) is substituted into equation (2) and the inner product part φ (x _i ) φ (x _j ) is replaced with kernel k (x _i , x _j ), equation (16) is obtained.

The bias term b is determined by substituting the feature vectors x _A1 and x _B1 for determining the classification margin shown in FIG. 6B into Equation (16).

式（１７）は、ｉ番目の画像の特徴ベクトルをｘ_ｉ，ｊ番目の画像の特徴ベクトルをｘ_ｊとし、ｘ_ｉとｘ_ｊの対応するｋ番目の要素を比較して、小さい方の要素を選択して累算し、そして、累算結果をα乗することを示している。 As described above, the learning unit 23 obtains the weight vector m and the bias term b that maximize F (λ), and constructs Expression (2) representing the classification surface.
In the learning unit 23, the kernel k (x _i , x _j ) in the equations (15) and (16) is expressed by the functional equation (17).

Equation (17) sets the feature vector of the i-th image to x _i and the feature vector of the j-th image to x _j and compares the corresponding k-th elements of x _i and x _j Is selected and accumulated, and indicates that the accumulated result is raised to α.

A calculation example in which the learning unit 23 obtains the similarity from the five-dimensional feature vector using Equation (17) for the kernel will be specifically described using FIG. FIG. 5A is the normalized five-dimensional feature vector x _i described with reference to FIG. 3B. The normalized feature vector x _j of another image for which the degree of similarity is to be obtained is assumed to be as shown in FIG. In FIG. 5C, the vector element in FIG. 5A is compared with the vector element in FIG. 5B to select the smaller vector element. First, the smaller vector x _j1 (0.00) is selected by comparing the first vector elements x _i1 and x _j1 of the feature vector x _i and x _j . The next vector element x _i2 is compared with x _j2 to select the smaller x _i2 (0.15). Compare the next vector element x _i3 to x _j3 to select the smaller x _i3 (0.07), compare x _i4 to x _j4 to select the smaller x _j4 (0.28), The smaller x _i5 (0.37) is selected by comparing x _i5 and x _j5 . The right end of FIG. 5C represents a value obtained by multiplying the calculated value (0.87) obtained by accumulating these values by α. FIG. 5C exemplifies the case where α = 1, so (0.87) is the value obtained by the equation (17) where α = 1.
Thus, the learning unit 23 uses two kernels of k (x _i , x _j ) from the five-dimensional feature vector x _i and the five-dimensional feature vector x _j using equation (17) for the kernel. The similarity between the two is being sought.

FIG. 5 (d) is a bar graph representing values of normalized elements of the feature vector of FIG. 5 (a). FIG. 5 (e) is a bar graph showing the values of the normalized elements of the feature vector of FIG. 5 (b). In FIG. 5 (f), FIG. 5 (d) and FIG. 5 (e) are superimposed and the smaller one is blackened. That is, the black part corresponds to {Σmin (x _ik , x _jk )}. As can be seen from this figure, {Σmin (x _ik , x _jk )} is the common part of feature vector x _i and feature vector x _j, and therefore represents the similarity (commonality) between feature vectors . The overlapping portion increases as the feature vector x _i and the feature vector x _j approximate.

The power parameter α of the equation (17) processed by the learning unit 23 does not evaluate the trivial information when the value of alpha is increased, and the function of emphasizing the trivial information and evaluation is performed when the value of α is decreased. Have.
Describing the case of similarity classification with respect to a certain face, if the power parameter α is reduced, the presence of the minor element which is a minor element in the face as a whole is emphasized and evaluated. For this reason, even if accurate classification can be made regarding the presence or absence of hokuro, there is a possibility that an image that is the same as the whole face but has no hokuro may be classified as non-similar.
The learning unit 23 obtains the degree of similarity between two feature vectors x _i and x _j using the equation (17) from the five-dimensional feature vector to the kernel. Then, feature vectors can be easily classified by performing dimensional expansion using the similarity between the feature vectors.
This is the end of the functional description of the learning unit 23.

Next, the operation of the information classification device 11 having the above configuration will be described. In order for the information classification device 11 to classify unknown images, it is necessary to set in advance in the information classification device 11 a classification surface for classifying unknown images. In order to set a classification plane in the information classification device 11, it is necessary to construct a classification plane using a learning image of known classification and a test image of known classification.
Hereinafter, details of processing until the classification plane is set in the information classification device 11 will be described using the functional configuration diagram shown in FIG. 2 and the flowchart diagram shown in FIG. The classification surface setting process shown in FIG. 7 is started by reading the non-linear SVM processing program from the ROM 16 and starting the learning function of the program.

  First, the user operates the keyboard of the input unit 12 to supply the learning image group and the test image group stored in the storage unit 15 to the acquisition unit 21 (step S1). For example, in the case of classifying the image of Mr. A and the other images from a large number of unspecified images, the image group of Mr. A and a large number of images of unspecified people are used as learning images, Are supplied from the storage unit 15 to the acquisition unit 21 as a test image.

Next, the user sets the range of the regularization parameter C indicating the tolerance of the classification error and the power parameter α of the kernel (step S2). In the following description, in order to facilitate understanding, the regularization parameter C has an initial value of C1, a maximum value of Cm, and a change magnification of Ct, where C1 = 0.001, Cm = 10, and Ct = 10. . The exponentiation parameter α has an initial value α1, a maximum value αn, and a change width αstep, and α1 = 0.1, αn = 20, and αstep = 0.2.
The learning unit 23 stores each parameter value input from the input unit 12 in the RAM 17. Alternatively, it may be stored in the ROM 16 or the storage unit 15 as an initial value held by the information classification device 11.

  The control unit 18 functioning as the acquisition unit 21 divides each of the supplied learning image and test image into 5 × 5 blocks using the Bof program stored in the ROM 16 and determines the color of each block Then, the feature vector is extracted from the image with the color frequency as a vector element, and stored in the RAM 17 (step S3). An example of the acquired feature vector corresponds to the second line in FIG.

  The normalization unit 22 normalizes each extracted feature vector with its own L2 norm (step S4).

  Next, the learning unit 23 constructs a classification plane for classifying the learning image while sequentially changing the combination of the regularization parameter C and the exponentiation parameter α set in step S2.

First, the learning unit 23 sets (C, α) = (C1, α1) = (0.001, 0.1) as a combination of the first regularization parameter C and the exponentiation parameter α (step S5, Step S6).
The learning unit 23 constructs a classification surface that classifies the normalized feature vector corresponding to the learning image within the allowable range of the classification error specified by the regularization parameter C (step S7). Specifically, the learning unit 23 uses Equation (17) as a kernel in a non-linear SVM processing program that constructs this classification plane, and generates similarity from normalized five-dimensional feature vectors of two learning images. The degree is determined and substituted into equation (15). When constructing a classification surface for classifying Mr. A out of 100 learning images, this process is performed ₁₀₀ C ₂ = 4950 times to maximize the classification margin while satisfying the constraint condition weight vector m And the bias term b, and formula (2) representing a classification surface for classifying Mr. A is constructed.

  Next, the test unit 24 classifies the feature vector corresponding to the test image whose known way is to be classified, using the constructed classification surface. Here, classification is performed using the classification surface constructed in step S7 corresponding to (C, α) = (0.001, 0.1). Specifically, the test unit 24 normalizes the test image to x in Formula (2) representing the classification plane in which the weight vector m and the bias term b established by the learning unit 23 are determined. Assign. Then, if y = 1, the feature vector is classified into group A (image of Mr. A), and if y = −1, the feature vector is classified into group B (not the image of Mr. A). This classification is performed 100 times if there are 100 test images. Then, since the test image is an image known how to classify, the test unit 24 identifies the proportion of the feature vector corresponding to the classified test image classified into the group to be classified. The rate (probability of being correctly classified) is determined, and stored in the RAM 17 in association with the classification surface (step S8).

Next, the learning unit 23 adds αstep (= 0.2) to α to set α = 0.3 in order to construct a classification surface for the next value of the power parameter α (step S9).
Next, the learning unit 23 determines whether the exponentiation parameter α is equal to or less than the maximum value α n (step S10). Here, since α = 0.3 (<20) (step S10: Yes), the process of step S7 and the process of step S8 are performed, and the second classification plane and the identification rate are stored in the RAM 17 .

This process is repeated, and when α> α n (α> 20) (step S10: No), the process proceeds to step S11, and Ct is changed to C = 0.001 to change the value of the regularization parameter C. 10) to set C = 0.01 (step S11).
Then, it is determined whether the regularization parameter C is equal to or less than the maximum value Cm (C ≦ 10) (step S12). Here, since C = 0.01, the process returns to step S6 (step S12: Yes). Since α is reset to α1 in step S6, the processing in steps S7 and S8 is performed as (C, α) = (0.01, 0.1), and the classification surface is identified and identified. The rate is stored in the RAM 17.

  When the processing to α = 0.1 to 20 is completed with respect to C = 0.01 (step S10: No), C is multiplied by Ct (step S11) to determine whether the regularization parameter C is the maximum value Cm or less It discriminates (step S12). This process is repeated until C> Cm is established, and when C> Cm (C> 10) is established (step S12: No), the process proceeds to step S13.

  The selection unit 25 sorts the identification rates obtained corresponding to the classification planes in the test unit 24 in descending order, and selects the classification plane associated with the highest identification rate. The setting unit 26 sets the selected classification plane to the nonlinear SVM implemented in the classification unit 27 of the information classification device 11 (step S13), and ends the process of setting the classification surface to the information classification device 11. Thus, the process of setting the classification plane in the information classification device 11 is completed.

Since the process of setting the classification plane in the information classification device 11 is completed, the user uses the information classification device 11 in which the above setting process is completed to classify an unknown image as to how to classify. The user starts the information classification function of the non-linear SVM processing program, designates the classification target image whose classification is unknown from the input unit 12, and supplies it to the acquisition unit 21. The acquisition unit 21 extracts a feature vector corresponding to an unknown classification target image, the normalization unit 22 normalizes the feature vector, and supplies the normalized feature vector to the classification unit 27. The classification unit 27 classifies the feature vector corresponding to the normalized unknown classification target image using the non-linear SVM for which the classification plane is set in step 13.
Specifically, the classification target image is classified into group A or group B by substituting the normalized feature vector corresponding to the classification target image into the functional expression (2) representing the selected classification surface. In this example, since the classification surface for classifying Mr. A is set, the image group of Mr. A and the image group of other people are input as a classification target image group, and the classification target image group is input from Mr. A The image of is classified as group A. This is the end of the description of the information classification device 11 classifying the classification target image.

本実施形態に用いるカーネルを用いた場合のべき乗計算の処理回数と、非特許文献１が開示するカーネルを用いた場合のべき乗計算の処理回数とを図８で比較する。本実施形態では、特徴ベクトルの要素数が５の場合で説明したが、人の画像を分類する場合には、目・鼻・口・耳などの色や形状や角度などの様々な情報をベクトル要素としている。分類精度を上げるためには、特徴ベクトルの要素の数を増やす必要がある。このように、特徴ベクトルの要素数が大きくなるほど、式（１７）をカーネルに使用する本実施形態に係る情報分類装置１１のべき乗計算処理の負荷は、従来技術に比べると相対的に著しく軽くなる。 As described above, the number of times of processing of the exponentiation calculation of Expression (17) used for the kernel described in this embodiment is only once even if the number of elements of the feature vector increases. On the other hand, in the kernel of equation (18) disclosed in Non-Patent Document 1, as the number K of elements of the feature vector increases, the number of times of processing of the power calculation also increases.

The number of times of processing of the exponentiation when using the kernel used in the present embodiment and the number of times of processing of the exponentiation when using the kernel disclosed in Non-Patent Document 1 are compared in FIG. In the present embodiment, the case where the number of elements of the feature vector is five has been described, but when classifying a human image, various information such as colors, shapes, angles, etc. of eyes, nose, mouth, ears etc. It is an element. In order to improve classification accuracy, it is necessary to increase the number of elements of the feature vector. As described above, as the number of elements of the feature vector increases, the load of the power calculation process of the information classification device 11 according to the present embodiment using Equation (17) for the kernel becomes relatively lighter compared to the prior art. .

  Further, since the equation (17) used for the kernel described in the present embodiment has a power term, the classification apparatus using the equation (17) is the same as the equation (18) disclosed in the prior art. By adjusting the power parameter α, it is possible to increase the options for constructing the classification surface. The information classification device using the equation (17) can reduce the processing load of the exponentiation calculation while having an effect by the exponentiation parameter. Therefore, by using the present embodiment, an information classification device having high classification accuracy and small processing load of kernel power calculation even if the number of elements of feature vectors increases is provided, and a setting method and program of the information classification device be able to.

Second Embodiment
The first embodiment has described the case where the setting function for setting the classification surface of the nonlinear SVM and the function for classifying an unknown image are implemented in an integrated device, but the present invention is not limited to this. The present invention is applicable to a setting device that constructs a classification surface using a power of similarity between feature vectors as a kernel, and an information classification device that classifies information using the classification surface constructed by the setting device. It can apply. As another embodiment, in the second embodiment, the case where the function of setting the classification surface of the nonlinear SVM and the function of classifying an unknown image are implemented in different devices will be described.
In the drawings, the same or corresponding parts are denoted by the same reference numerals.

A functional configuration diagram of the setting device 11A for setting the classification surface of the non-linear SVM according to the embodiment of the present invention is shown in FIG. A functional block diagram of an information classification device 11B according to an embodiment of the present invention, the classification surface of which is set by the setting device 11A, is shown in FIG.
The hardware configuration of the setting device 11A and the information identification device 11B may be the same as the configuration described using FIG. 1 in the first embodiment.

The functional configuration of the setting device 11A according to the present embodiment will be described with reference to FIG. 9, focusing on the difference from the first embodiment.
The difference between the information classification device 11 according to the first embodiment and the setting device 11A according to the second embodiment is that the setting device 11A does not have the classification unit 27. That is, the setting device 11A according to the second embodiment selects the classification surface having the highest identification rate, and sets the selected classification surface in the information classification device 11B shown in FIG. Along with this, the input unit 12 of the setting device 11A does not issue an instruction to input a classification target image whose method of classification is unknown.

First, a known learning image and a known test image are supplied to the acquiring unit 21 in order to perform processing of constructing a classification plane in the setting device 11A. The acquisition unit 21 extracts a feature vector from the learning image group and the test image group using the Bof program, and supplies the feature vector to the normalization unit 22. The normalization unit 22 normalizes the feature vector with the L2 norm.
The learning unit 23 constructs a classification plane for classifying the normalized feature vector of the learning image so as to maximize the classification margin, using the equation (17) for the kernel of the nonlinear SVM registered. . The learning unit 23 constructs this classification plane for each combination of the regularization parameter C and the power parameter α. The test unit 24 classifies a feature vector corresponding to a known test image using the classification surface constructed by the learning unit 23. Then, the test unit 24 obtains an identification rate for each of the plurality of classification planes constructed by the learning unit 23.
The selection unit 25 selects the classification plane having the highest discrimination rate among the discrimination rates obtained for each classification plane in the test unit 24. The setting unit 26 stores, in the storage unit 15, the classification surface having the highest discrimination rate selected by the selection unit 25. Alternatively, the input / output I / F unit 14 outputs information on the selected classification surface to a recording medium such as a USB memory. Alternatively, by connecting the input / output I / F unit 14 of the setting device 11A and the input / output I / F unit 14 of the information classification device 11B shown in FIG. 10, the classification surface selected from the setting device 11A can be used as the information classification device 11B. It may be set to the nonlinear SVM implemented in.

Next, the functional configuration of the information classification device 11B according to the present embodiment will be described with reference to FIG. 10, focusing on the difference from the first embodiment.
The difference between the information classification device 11 according to the first embodiment and the information classification device 11B according to the second embodiment is that the information classification device 11B does not have the learning unit 23, the test unit 24, the selection unit 25, and the setting unit 26. is there. That is, the information classification device 11B according to the second embodiment is an information classification device that classifies an unknown image in which the classification plane selected by the setting device 11A is set.

First, the user operates the keyboard of the input unit 12 to supply the acquisition unit 21 with a classification target image whose classification is unknown stored in the storage unit 15. The acquisition unit 21 extracts a feature vector from the classification target image using the Bof program, and supplies the feature vector to the normalization unit 22. The normalization unit 22 normalizes the feature vector with the L2 norm, and supplies the normalized feature vector to the classification unit 27. The classification unit 27 classifies the feature vector corresponding to the classification target image using the non-linear SVM in which the classification plane is set.
Specifically, the classification unit 27 substitutes the normalized feature vector corresponding to the classification target image into the functional expression (2) representing the set classification surface, and sets the classification target image to the group A or the group B. Categorize. In this example, since the classification surface for classifying Mr. A is set, the image group of Mr. A and the image group of other people are input as a classification target image group, and the classification target image group is input from Mr. A The image of is classified as group A.

  The process until the classification surface to be set in the classification unit 27 of the information classification device 11B is selected in the setting device 11A is the same as the flowchart shown in FIG.

  As described above, in the second embodiment, the function of setting the classification plane and the function of classifying unknown information are implemented in separate apparatuses. Therefore, the processing load of the information classification device 11B that does not require the classification surface construction process can be reduced. Further, since it is not necessary to store the information for the learning image and the test image in the storage unit 15 or the RAM 17 of the information classification device 11B, the memory amount can be reduced. Thus, the information classification device 11B can be manufactured inexpensively by reducing the processing load of the information classification device 11B.

In the second embodiment, the classification surface selected by the setting device 11A is set in the information classification device 11B via a storage medium or the like. However, the classification surface selected by the setting device 11A is on the communication network such as the Internet It may be stored in a storage unit in the server device of the above, and may be downloaded to the information classification device 11B.
In the second embodiment, the unknown classification target image is described as being stored in the storage unit 15. However, the unknown classification target image may be supplied via a communication network such as the Internet.

(Modification)
In the above embodiment, the case where equation (17) for finding the similarity between feature vectors is used as a kernel for finding the relationship between feature vectors has been described. However, the equation for obtaining the degree of similarity is not limited to this. For example, equation (19) for determining the difference between feature vectors can also be used.

式（２０）で求めた計算値は、特徴ベクトルｘｉと特徴ベクトルｘ_ｊとの類似度が高くなるほど１に近づく。 Further, as a formula for finding the degree of similarity, it is also possible to use a formula for finding the ratio of feature vectors, for example, formula (20).

The calculated value obtained by equation (20) approaches 1 as the degree of similarity between the feature vector xi and the feature vector x _j increases.

  As described above, any other equation may be used as a kernel as long as the equation can be used to determine the similarity between feature vectors.

  Although the information classification apparatus according to the present invention is used as an image classification apparatus in the description of the embodiment, the application of the present invention can be used for classification of various information in addition to image classification. . For example, music can be classified as waltz or not, or major and minor. The present invention can also be applied to other applications where feature vectors are extracted from information and classified using an information classification device.

  Further, in the present embodiment, a flowchart for performing the learning step and the test step every time one combination of the regularization parameter C and the exponentiation parameter α is set is described using FIG. 7. However, after completing the learning process for all combinations of the regularization parameter C and the power parameter α, the test process may be performed on all of the constructed classification planes.

Further, the values of the parameters used in the present embodiment are examples for describing the invention, and the parameter values used in the embodiment do not affect the scope of the invention.
In the above embodiment, although the regularization parameter C and the exponentiation parameter α are changed according to a certain rule, a plurality of parameter values whose change rules are not constant are stored in the storage unit 15 , May be read out sequentially and used. For example, the regularization parameter C and the exponentiation parameter α may be generated randomly. Alternatively, one of the regularization parameter C and the exponentiation parameter α may be fixed and used.

Further, various functions realized by the control unit 18 according to the present invention can be realized not only by a dedicated system but also by using a normal computer system. For example, by storing a program in a computer-readable recording medium such as a flexible disk, a compact disc read only memory (CD-ROM), a USB memory, etc. and storing the program in a computer, the functions of the above-described portions can be realized. You may configure a computer.
Further, the program may be stored in a server apparatus connected to a communication network such as the Internet so that, for example, the computer can download the program.

  Although the preferred embodiments of the present invention have been described above, the present invention is not limited to the specific embodiments, and the present invention includes the invention described in the claims and the equivalents thereof. included. The invention described in the claims at the beginning of the present application is appended below.

(Supplementary Note 1)
An input unit for inputting a plurality of learning information whose classification is known;
Acquisition means for acquiring feature vectors corresponding to each of the learning information;
Learning means for constructing a classification surface for classifying the learning information on the basis of the power of similarity between feature vectors;
Setting means for setting the classification surface constructed by the learning means in the information classification device;
Configuration device comprising:

(Supplementary Note 2)
The input means further includes a plurality of test information of known classification, a plurality of regularization parameters indicating tolerance of misclassification of the learning information, and a plurality of power parameters for raising the similarity between feature vectors. And enter
The acquisition means acquires feature vectors respectively corresponding to the test information,
The learning means constructs a plurality of classification planes for each combination of the regularization parameter and the exponentiation parameter,
Furthermore, one of a plurality of classification planes constructed by the learning means is selected, a feature vector corresponding to the test information prepared by the acquiring means is classified, and a feature vector corresponding to the test information is classified Determining the identification rate representing the proportion classified into the known classification state to be performed, and further updating the selected classification surface sequentially and performing the test a plurality of times;
Selection means for selecting a classification surface having the highest identification rate obtained by the test means;
The setting device according to claim 1, further comprising:

(Supplementary Note 3)
The acquisition means acquires a feature vector which quantifies features of information by Bof (Bag of Features) from the information for learning and the information for test.
And a normalization means for normalizing the acquired feature vector with L2 norm,
The learning means constructs a plurality of classification planes for classifying the feature vectors normalized by the normalization means.
The setting device according to appendix 2, characterized in that

(Supplementary Note 4)
An information classification unit that classifies unknown information different from the learning information and the test information according to the classification surface constructed by the setting device according to any one of appendices 1 to 3;
An information classification device comprising:

(Supplementary Note 5)
An input step of inputting a plurality of learning information whose classification is known;
An acquisition step of acquiring feature vectors corresponding to each of the learning information;
A learning step of constructing a classification surface for classifying the information for learning based on a power value of similarity between feature vectors;
A setting step of setting the classification surface constructed in the learning step in the information classification device;
Classification plane setting method of setting device including.

(Supplementary Note 6)
An information classification step of classifying unknown information different from the learning information and the test information according to the classification surface set by the classification surface setting method of the setting device as set forth in appendix 5;
And an information classification method of the information classification device.

(Appendix 7)
Computer,
Input means for inputting a plurality of learning information of known classification;
Acquisition means for acquiring feature vectors corresponding to each of the learning information;
Learning means for constructing a classification surface for classifying the learning information based on a power of similarity between feature vectors;
Setting means for setting the classification plane constructed by the learning means in the information classification device;
Program to function as.

(Supplementary Note 8)
Computer,
An information classification unit that classifies unknown information different from the learning information and the test information according to the classification surface set by the execution of the program described in the supplementary note 7;
Program to function as.

11: Information classification device, 11A: Setting device, 11B: Information classification device, 12: Input unit, 13: Display unit, 14: Input / output I / F, 15: Storage unit, 16: ROM, 17: RAM, 18: Control part, 21 ... acquisition part, 22 ... normalization part, 23 ... learning part, 24 ... test part, 25 ... selection part, 26 ... setting part, 27 ... classification part

Claims (7)

A plurality of learning information having a known classification, a plurality of test information having a known classification, a plurality of regularization parameters indicating tolerance of misclassification of the learning information, and a power of similarity between feature vectors Input means for inputting a plurality of power parameters ,
Acquisition means for acquiring feature vectors respectively corresponding to the learning information and the test information ;
Learning means for constructing a plurality of classification planes for classifying the learning information for each combination of the regularization parameter and the exponentiation parameter based on the power of the similarity between feature vectors;
One of a plurality of classification planes constructed by the learning means is selected, a feature vector corresponding to the test information prepared by the acquiring means is classified, and a feature vector corresponding to the test information is classified Testing means for determining a classification rate representing the rate of classification into known classification states, and further updating the selected classification surface sequentially and performing the process multiple times;
Selection means for selecting a classification surface having the highest identification rate obtained by the test means;
Setting means for setting the classification surface selected by the selection means in the information classification device;
Configuration device comprising: The acquisition means acquires a feature vector which quantifies features of information by Bof (Bag of Features) from the information for learning and the information for test.
And a normalization means for normalizing the acquired feature vector with L2 norm,
The learning means constructs a plurality of classification planes for classifying the feature vectors normalized by the normalization means.
Setting device according to claim 1, characterized in that. An information classification unit that classifies unknown information different from the learning information and the test information according to the classification surface constructed by the setting device according to claim 1 or 2 .
An information classification device comprising: A plurality of learning information having a known classification, a plurality of test information having a known classification, a plurality of regularization parameters indicating tolerance of misclassification of the learning information, and a power of similarity between feature vectors An input step of inputting a plurality of power parameters ;
An acquiring step of acquiring feature vectors respectively corresponding to the learning information and the test information ;
A learning step of constructing a plurality of classification planes for classifying the learning information for each combination of the regularization parameter and the exponentiation parameter based on the power of the similarity between feature vectors;
One of a plurality of classification planes constructed in the learning step is selected, a feature vector corresponding to the test information generated in the acquisition step is classified, and a feature vector corresponding to the test information is classified A test process in which a classification rate representing a ratio classified into a known classification state is determined, and a classification surface to be selected is sequentially updated and performed a plurality of times;
A selection step of selecting a classification surface having the highest identification rate obtained in the test step;
A setting step of setting the classification surface selected by the selection step in the information classification device;
Classification plane setting method of setting device including. An information classification step of classifying unknown information different from the learning information and the test information according to the classification surface set by the classification surface setting method of the setting device according to claim 4 ;
And an information classification method of the information classification device. Computer,
A plurality of learning information having a known classification, a plurality of test information having a known classification, a plurality of regularization parameters indicating tolerance of misclassification of the learning information, and a power of similarity between feature vectors Input means for inputting a plurality of power parameters ,
Acquisition means for acquiring feature vectors respectively corresponding to the learning information and the test information ;
Learning means for constructing a plurality of classification planes for classifying the learning information for each combination of the regularization parameter and the exponentiation parameter based on the power of the similarity between feature vectors;
One of a plurality of classification planes constructed by the learning means is selected, a feature vector corresponding to the test information prepared by the acquiring means is classified, and a feature vector corresponding to the test information is classified A test means for determining a classification rate representing the proportion classified into the known classification state, and further updating the selected classification surface sequentially and performing the process multiple times,
Selection means for selecting a classification surface having the highest discrimination rate obtained by the test means,
Setting means for setting the classification surface selected by the selection means in the information classification device;
Program to function as. Computer,
An information classification unit that classifies unknown information different from the learning information and the test information according to the classification surface set by the execution of the program according to claim 6 .
Program to function as.

Images