JP5633424B2 - Program and information processing system - Google Patents

Description

  The present invention relates to a program and an information processing system.

  In recent years, an image automatic annotation technique for automatically assigning keywords (signs and labels) for retrieval to images has become one important technique for image retrieval. This technique makes it unnecessary for a person to add a keyword to an image.

  In Non-Patent Document 1, a local feature amount is extracted from an image, the distribution of the entire image is calculated as an image feature, and a label is generated by a probabilistic binary support vector machine provided for each label. A technique is disclosed in which a posterior probability is calculated and a label corresponding to the highest posterior probability is assigned to an unknown image.

K.S.Goh, "Using One-Classand Two-Class SVMs for Multiclass Image Annotation", IEEE Trans.OnKnowledge and Data Engineering, Vol.17, No.10, Oct.2005

  An object of the present invention is to increase the classification accuracy of an image as compared with the prior art.

  The invention according to claim 1 is a program, an image receiving unit that receives an image that is a target for specifying at least one of a plurality of classifications, and the image receiving unit receives each of the plurality of classifications. Based on the feature quantity of the image and the feature quantities of a plurality of images each belonging to at least one of the plurality of classifications, it is determined whether the image received by the image receiving means belongs to the classification. A discrimination reference value specifying means for specifying a discrimination reference value serving as a reference; each of the plurality of classifications is specified based on a plurality of images belonging to at least one of the plurality of classifications; At least one of the possibility that an image belongs to another category or the possibility that an image that does not belong to the category does not belong to the other category. Based on the correlation information representing the classification and the discrimination reference value for the classification and the other classification, the classification affiliation specifying the value representing the probability that the image received by the image receiving means belongs to the classification Computer as output means for outputting information indicating at least one classification to which the image received by the image receiving means, specified based on the value specified by the possibility specifying means, the classification attribute possibility specifying means, belongs Is supposed to function.

  Invention of Claim 2 is the program of Claim 1, Comprising: The said classification genus possibility determination means is the said correlation information about each classification | category different from the said classification | category about each said classification | category, and Based on the combination of the discriminant reference values, a value representing the possibility that the image received by the image receiving means belongs to the classification is specified.

  The invention according to claim 3 is an information processing system, wherein an image receiving unit that receives an image for specifying at least one of a plurality of classifications, and the image reception for each of the plurality of classifications Whether the image received by the image receiving means belongs to the category based on the feature amount of the image received by the means and the feature amounts of a plurality of images each belonging to at least one of the plurality of classifications A discrimination reference value specifying means for specifying a discrimination reference value serving as a reference for discriminating, and for each of the plurality of classifications, each is specified based on a plurality of images belonging to at least one of the plurality of classifications, There is little possibility that an image belonging to the category belongs to another category, or an image that does not belong to the category does not belong to the other category. Based on the correlation information representing at least one and the discrimination reference value for the classification and the other classification, a value representing the possibility that the image received by the image receiving means belongs to the classification is specified. Output that outputs information indicating at least one classification to which an image received by the image receiving unit is specified, which is specified based on the value specified by the classification attribute possibility determining unit and the classification attribute possibility determining unit Means.

  According to the first and third aspects of the invention, the image classification accuracy is higher than in the case of not having the configuration of the present invention.

  According to the second aspect of the present invention, the possibility that an image belongs to a certain class is specified based on the correlation between all classes different from the class.

It is a figure which shows an example of the hardware constitutions of the information processing apparatus which concerns on one Embodiment of this invention. It is a functional block diagram which shows an example of the function implement | achieved by the information processing apparatus which concerns on one Embodiment of this invention. It is a figure which shows an example of learning model information. It is a figure which shows an example of 1st correlation matrix information. It is a figure which shows an example of 2nd correlation matrix information. It is a figure which shows an example of probability model information. It is a flowchart which shows an example of the flow of the process performed with the information processing apparatus which concerns on this embodiment. It is a figure which shows an example of the probability model information in a comparative example. It is a figure which shows an example of 1st correlation matrix information. It is a figure which shows an example of 2nd correlation matrix information. It is a figure which shows an example of the comparison result of the recognition rate of a label | marker. It is a figure which shows an example of the sigmoid function of the label | marker shape optimized by the discrimination | determination reference value with respect to the marker motorbike. It is a figure which shows an example of the sigmoid function of the label | marker shape optimized by the discrimination | determination reference value with respect to the marker motorbike. It is a figure which shows an example of the comparison result of the recognition rate of a label | marker. It is a figure which shows an example of the sigmoid function of the label | marker cat optimized with the discrimination | determination reference value with respect to the label | marker person. It is a figure which shows an example of the sigmoid function of the label | marker cat optimized with the discrimination | determination reference value with respect to the label | marker person. It is a figure which shows one specific example of the parameter tilde A contained in probability model information.

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

  FIG. 1 is a diagram illustrating an example of a hardware configuration of an information processing apparatus 10 that functions as an information processing system according to the present embodiment. As illustrated in FIG. 1, the information processing apparatus 10 according to the present embodiment includes, for example, a control unit 12, a storage unit 14, and a user interface (UI) unit 16. These elements are connected via a bus or the like. The control unit 12 is a program control device such as a CPU, and operates according to a program installed in the information processing apparatus 10. The storage unit 14 is a storage element such as a ROM or a RAM, a hard disk drive, or the like. The storage unit 14 stores a program executed by the control unit 12. The storage unit 14 also operates as a work memory for the control unit 12. The UI unit 16 is a display, a microphone, a mouse, a keyboard, and the like, and outputs the content of the operation performed by the user and the voice input by the user to the control unit 12. In addition, the UI unit 16 displays and outputs information according to an instruction input from the control unit 12.

  FIG. 2 is a functional block diagram illustrating an example of functions realized by the information processing apparatus 10 according to the present embodiment. As illustrated in FIG. 2, in the present embodiment, the information processing apparatus 10 includes, for example, an image reception unit 20, a learning model information storage unit 22, a correlation matrix information storage unit 24, a probability model information storage unit 26, and feature amount extraction. The unit 28, the discrimination reference value calculation unit 30, the discrimination reference value weighting unit 32, the affiliation possibility calculation unit 34, the classification determination unit 36, the output unit 38, and the learning unit 40 function. The learning model information storage unit 22, the correlation matrix information storage unit 24, and the probability model information storage unit 26 are realized mainly by the storage unit 14. Other elements are realized mainly by the control unit 12. The learning unit 40 includes a discriminator such as a support vector machine, for example.

  These elements are realized by executing a program installed in the information processing apparatus 10 that is a computer by the control unit 12 of the information processing apparatus 10. This program is supplied to the information processing apparatus 10 via a computer-readable information recording medium such as a CD-ROM or DVD-ROM, or via communication means such as the Internet.

  In the present embodiment, for example, at least one of L types of labels (labels) determined in advance is associated with the image received by the image receiving unit 20. In the present embodiment, a label number j (j = 1,..., L) is assigned to each of the L types of labels. In the present embodiment, there is a one-to-one correspondence between the sign and the image classification. That is, an image belonging to a certain category is associated with a label corresponding to that category. An image may belong to a plurality of categories. In that case, a plurality of signs are associated with the image.

In the present embodiment, learning model information 50 illustrated in FIG. 3 is stored in the learning model information storage unit 22 in advance. In the present embodiment, the learning model information 50 is generated by the information processing apparatus 10 learning a plurality of images (learning images) to be learned. The learning model information 50 includes the above-described label number, support vector information, and parameter information. The learning model information 50 is information representing a learning model of a binary support vector machine that identifies, for example, a feature quantity corresponding to a sign and a feature quantity corresponding to another sign. The support vector information and the parameter information are information serving as a basis for calculating a discrimination reference value described later. Further, as shown in FIG. 3, the values of the support vector information associated with the label number i are (x _i1 , x _i2 ,..., X _iNi ), and the parameter information values are (a _i1 , a _{i2 ,. ..} , A _iNi ) and b _i . The generation process of the learning model information 50 based on the learning image will be described later.

Further, in the present embodiment, first correlation matrix information 52-1 exemplified in FIG. 4 and second correlation matrix information 52-2 exemplified in FIG. 5 are stored in the correlation matrix information storage unit 24 in advance. Has been. The first correlation matrix information 52-1 and the second correlation matrix information 52-2 are generated by the information processing apparatus 10 learning a plurality of learning images. These learning images may be the same as or different from the learning images used when the learning model information 50 is generated. For example, the conditional probability P in which the sign of the sign number i is also associated with the learning image in which the sign of the sign number j is associated with the value R _ij of the i row and j column of the first correlation matrix information 52-1. Absolute value R _ij = | P (C _i ) −P (C _i | C) of the difference between (C _i | C _j ) and the probability P (C _i ) that the sign of the sign number i is associated with the learning image _j ) | is set. When the sign of the sign number j is associated with the learning image and the sign of the sign number i is also associated with the learning image, the value of P (C _i | C _j ) is larger than the case where the sign is not. growing. On the other hand, when the sign of the sign number j is not associated with the learning image and the sign of the sign number i is not associated with the learning image, P (C _i | C _j ) The value becomes smaller. In any case, if there is a correlation as to whether or not the label with the label number i and the label with the label number j are associated with the learning image, | P (C _i ) −P (C _i | C _j ) | Increases. On the other hand, when there is no correlation, the value of | P (C _i ) −P (C _i | C _j ) | If independent, the value of | P (C _i ) −P (C _i | C _j ) | is zero.

On the other hand, the value R ′ _ij in the i-th row and j-th column of the second correlation matrix information 52-2 is, for example, a condition in which the sign of the sign number i is not associated with the learning image that is not associated with the sign of the sign number j. The absolute value R ′ _ij = | P (C ′ _i ) of the difference between the assigned probability P (C ′ _i | C ′ _j ) and the probability P (C′i) that the sign of the sign number i is not associated with the learning image ) -P (C ′ _i | C ′ _j ) | is set.

In the present embodiment, probability model information 54 illustrated in FIG. 6 is stored in advance in the probability model information storage unit 26. In the present embodiment, the probability model information 54 is generated by the information processing apparatus 10 learning a plurality of learning images. These learning images may be the same as or different from the learning images used when the probability model information 54 is generated. Probability model information 54, the L rows and L parameters tilde A and element (i row assigned to each line is a matrix (A _ij) values of i-th row and the j which is arranged ~ tilde A _{ij (upper)} column the value of this element to be associated is contains parameters tilde B is a tilde B _{i (B i} of ~ upward are _arranged)). In the present embodiment, as shown in FIG. 6, the parameter in the i-th row in the probability model information 54 is a parameter corresponding to the marker with the marker number i. In the present embodiment, the probability model information 54 is a parameter vector of a sigmoid function, for example. The generation process of the probability model information 54 based on the learning image will be described later.

  Here, an example of the flow of processing performed by the information processing apparatus 10 according to the present embodiment will be described with reference to the flowchart illustrated in FIG.

  First, the image receiving unit 20 receives an image to be classified (that is, an image to which at least one of the above-described L types of signs is associated), and outputs a pixel matrix of this image (S101). ). Hereinafter, the image received in the process shown in S101 is referred to as a received image. Note that general preprocessing (for example, movement, color correction, transformation, format conversion, noise deletion, etc.) may be performed in the processing shown in S101.

  Then, the feature amount extraction unit 28 specifies, for example, the feature amount of the received image (S102). In the present processing example, the feature amount extraction unit 28, for example, the RGB value, normalized-RGB value, HSV value, LAB value, robustHue feature amount (van de Weijer, C. Schmid, “Coloring”) of each pixel included in the received image. Local Feature Extraction "(Refer to ECCV 2006) Gabor feature value, DCT feature value, SIFT feature value, GIST feature value, etc. are specified. Then, the feature quantity extraction unit 28 quantizes the extracted feature quantity in advance based on a codebook generated by K-Means clustering from the learning corpus. Then, the feature amount extraction unit 28 specifies the histogram of the entire received image of the quantized feature amount as the feature amount vector x in the process shown in S102.

Then, the discrimination reference value calculation unit 30 determines that each of the received images is based on the feature amount of the received image specified in the process shown in S102 and the learning model information 50 stored in advance in the learning model information storage unit 22. A discrimination reference value, which is a reference value for discriminating whether or not the image belongs to the classification (whether or not the sign is associated with the received image) is calculated (S103). In the present embodiment, the discrimination reference value calculation unit 30 calculates a discrimination reference value for each classification (for each marker), for example. For example, the discrimination reference value calculation unit 30 calculates the discrimination reference value F _i (i = 1,..., L) associated with the label with the label number i by calculating the following equation. In the following equation, K is a kernel function (for example, Gaussian kernel), and the bold letter x is a feature vector of the received image specified by the processing shown in S102. x _in is a value of support vector information included _in the learning model information 50. The values of a _in and b _i are the values of the parameter information included in the learning model information 50. The discrimination reference value F _i takes a large value for an image belonging to the classification corresponding to the sign, and takes a small value for other images. The discrimination reference value is, for example, an output of a decision function of a binary support vector machine that identifies a corresponding classification and another classification.

The discrimination reference value weighting unit 32 calculates the correlation matrix information 52 for the discrimination reference value F _i (i = 1,..., L) associated with the label with the label number i calculated in the process shown in S103. Is weighted based on (S104). Specifically, the discrimination reference value weighting unit 32 calculates, for example, L weighted discrimination reference values associated with the label for each label. Hereinafter, L weighted discrimination reference values associated with the label with the label number i are represented by F ′ _ij (j = 1,..., L).

For example, the discrimination reference value weighting unit 32 sets the first correlation matrix information when the value of the discrimination reference value F _j associated with the label with the label number j (j = 1,..., L) is greater than zero. It is checked whether the value R _ji included in 52-1 is larger than a predetermined threshold value. Then, the discrimination reference value weighting unit 32 sets the value of the weight w _ij to 1 when the value R _ji is larger than the threshold value, and sets the value of the weight w _ij to 0 otherwise. The value R ′ _ji included in the second correlation matrix information 52-2 when the value of the discrimination reference value F _j associated with the label with the label number j (j = 1,..., L) is smaller than zero. Is greater than a predetermined threshold value. Then, the discrimination reference value weighting unit 32 sets the value of the weight w _ij to 1 when the value R ′ _ji is larger than the threshold value, and sets the value of the weight w _ij to 0 when not. Then, the discrimination reference value weighting unit 32 calculates the value of F ′ _ij (j = 1,..., L) according to the mathematical formula F ′ _ij = F _j × w _ij . By repeatedly executing the above process for each i (i = 1,..., L), the value of F ′ _ij (i = 1,..., L, j = 1,..., L) Is calculated.

Then, the affiliation possibility calculation unit 34 calculates the following expression based on the value of F ′ _ij and the probability model information 54, so that each i (i = 1,..., L) is calculated. calculates a value of attribution possible values tilde _{p i (p i} to ~ upward are arranged) (S105). In the following equation, T represents a feature amount. The values of tilde A _ij and tilde B _i indicate values included in the probability model information 54 described above. The tilde p _i is a posterior probability corresponding to the label with the label number i.

Then, the classification determination unit 36 determines the classification to which the received image belongs (that is, the label associated with the received image) based on the value of the tilde p _i (i = 1,..., L) (S106). In the present embodiment, the classification determination unit 36 identifies a label number whose tilde p _i (i = 1,..., L) is greater than or equal to a predetermined threshold value, for example. Then, the classification determination unit 36 specifies a number of label numbers equal to or less than a predetermined number in order from the identified label number in descending order of the value of the tilde p _i (i = 1,..., L). . The classification determination unit 36 specifies the classification corresponding to the label number specified in this way as the classification to which the received image belongs.

  Then, the output unit 38 associates at least one sign corresponding to the classification determined in the process shown in S106 with the received image, outputs the sign to the storage unit 14, and outputs the sign number to the UI unit 16 such as a display. (S107). Thus, according to this processing example, the classification to which the received image belongs is determined, the corresponding sign is associated with the received image, and the sign number is displayed and output on a display or the like.

  Here, an example of the generation process of the learning model information 50 by the information processing apparatus 10 will be described.

First, the image receiving unit 20 receives a plurality (for example, N) of images (learning images) to be learned. Each learning image is associated with at least one sign corresponding to the classification to which the learning image belongs. Then, the feature amount extraction unit 28 extracts the feature amount of each learning image. Specifically, the feature quantity extraction unit 28, for example, includes RGB values, normalized-RGB values, HSV values, LAB values, robustHue feature values (van de Weijer, C. Schmid, “Coloring”) of each pixel included in the learning image. Local Feature Extraction "(see ECCV 2006), Gabor feature, DCT feature, SIFT feature, GIST feature, etc. Then, the feature quantity extraction unit 28 quantizes the extracted feature quantity in advance based on a codebook generated by K-Means clustering from the learning corpus. Then, the feature amount extraction unit 28 specifies the histogram of the entire received image of the quantized feature amount as the feature amount vector of the learning image. Hereinafter, a feature vector identified from the nth learning image (n = 1,..., N) is represented by _xn .

And the learning part 40 performs the specific process of support vector information and parameter information about each identification number j (j = 1, ..., L). Here, specific processing of support vector information and parameter information of the identification number j is, for example, the learning image of the n (n = 1, ···, N) features and the vector x _n of the learning image of the n A value t _n indicating whether or not it is associated with the sign corresponding to the identification number j (t _n = + 1 if associated with the sign corresponding to the identification number j. Otherwise, t _n = 0. ) (This process generates N sets of vectors and combinations of values (x ₁ , t ₁ ),..., (X _N , t _N )), and these N sets of vectors and Processing to learn combinations of values by support vector machine (SVM) (for example, N. Cristianini and J. Shawe-Taylor, "An Introduction to Support Vector Machines and Other Kernel-based Learning Methods", Chapter 6. Cambridge University Press 2000 Under optimization criteria, 1-v-Others Process of learning the parameters of the binary SVM), the process of identifying the support vector information and parameter information corresponding to the identification number j, is realized by a run in sequence. In the present embodiment, the learning section 40 learns the feature quantity vector x _n which is associated with t n _{= +} 1 as a positive example, the feature vector x _n which is associated with t n _{= 0} as a negative example To do. Then, the learning unit 40 outputs the learning model information 50 in which the corresponding identification number is associated with the support vector information and the parameter information specified as described above, to the learning model information storage unit 22.

  In this way, the learning model information 50 is stored in the learning model information storage unit 22.

  Next, an example of processing for generating the probability model information 54 by the information processing apparatus 10 will be described.

First, the image receiving unit 20 receives a plurality of images (learning images) to be learned. These learning images may be the same as or different from the learning images used for the generation process of the learning model information 50. Here, it is assumed that the number of learning images used for learning the sign of the sign number i is M _i . Each learning image is associated with at least one sign corresponding to the classification to which the learning image belongs. Then, the information processing apparatus 10 identifies the weighted determination reference value (F _ij ⁿ ) ′ for each learning image by the same processing as the processing shown in S102 to S104 described above. Then, the learning unit 40 treats the weighted discrimination reference value of the learning image associated with the learning target sign (for example, the sign with the sign number i) as a positive example (t _i ⁿ = + 1), and performs learning. A weighted discrimination reference value of a learning image that is not associated with a target sign (for example, a sign with a sign number i) is handled as a negative example (t _i ⁿ = 0), and is expressed by the following equation: The parameter tilde A _ij and tilde B _i that give the maximum likelihood are calculated by Newton's method using backtracking linear search (Nocedal, J. and SJWright: “Numerical Optimization” Algorithm 6.2. New York , NY: Springer-Verlag, 1999.). In the following equation, the value of (F _ij ⁿ ) ′ represents a weighted discrimination reference value for the n-th learning image (n = 1,..., M _i ) used for learning the marker with the marker number i. ing. T represents a feature amount.

Then, the learning unit 40 outputs the probability model information 54 including the values of the tilde A _ij and the tilde B _i specified as described above to the probability model information storage unit 26.

  In this way, the probability model information 54 is stored in the probability model information storage unit 26.

  Here, a comparative example of the present embodiment will be described.

  The information processing apparatus 10 according to this comparative example is different from the configuration illustrated in FIG. 2 in that the determination reference value weighting unit 32 is not included.

In this comparative example, the probability model information 54 illustrated in FIG. 8 is stored in the probability model information storage unit 26. In the probability model information 54, for example, two parameters (for example, a parameter A _i and a parameter B _i for the identification number _i ) are associated with each identification number.

And in this comparative example, the learning part 40 is the case where the label | marker (For example, the label | marker of the label | marker number i) with which the kth learning image becomes learning object is linked | related about each of several image used as learning object. The discrimination reference value F _k is handled as a positive example (t _k = + 1), and the discrimination reference value F _k is handled as a negative example when a label to be learned (for example, a label with the label number i) is not associated ( After setting t _k = 0), parameters A _i and B _i maximizing the likelihood expressed by the following equation are obtained by Newton's method using backtracking linear search (Nocedal , J. and SJWright: “Numerical Optimization” Algorithm 6.2. New York, NY: Springer-Verlag, 1999.). In the formula, the value of F _k denotes the value of the discriminant reference values F _k for the k-th learning image used for learning of the label of the labeled number i. The values of A and B indicate the values of parameters A and B corresponding to the label number i, respectively.

Then, in this comparative example, the attribute possibility calculation unit 34, based on the value of the discrimination reference value F _i (i = 1,..., L) for the received image and the probability model information 54, by calculating the following equation, each i (i = 1, ···, L) for, it calculates a value of attribution potential value _{p i} (S105). In the following formulas, a posterior probability p _i corresponds to the label of the labeled number i. The values of A and B indicate the values of parameters A and B corresponding to the label number i, respectively.

  The information processing apparatus 10 according to the present embodiment is different from at least the information processing apparatus 10 according to the comparative example in that the correlation matrix information 52 is used to calculate the value of the possibility of belonging.

  Here, an example of a comparison result between the case where the information processing apparatus 10 according to the present embodiment calculates the weighted determination reference value based on the correlation matrix information 52 and the case where the weighted determination reference value is not performed will be described. Here, the first correlation matrix information 52-1 illustrated in FIG. 9 and the second correlation matrix information 52-2 illustrated in FIG. 10 are used.

In the following description, for example, the value of the weighted discriminating reference value associated with the sign “sheep” is F _sheep , the posterior probability corresponding to the sign “sheep” is P _Sheep , and the value of t described above corresponding to the sign “sheep” is t _Sheep . Represent. In addition, it is set as the same expression also about labels other than label | marker sheep.

  First, an example of a comparison result between when the weighted discrimination reference value is calculated based on the first correlation matrix information 52-1 and when not calculated will be described. FIG. 11 shows a comparison result of the recognition rate of the sign when the weighted discrimination reference value is calculated using the first correlation matrix information 52-1 and when it is not calculated. In the comparison performed in FIG. 11, an image having a discrimination reference value larger than 0 was used. In FIG. 11, circles indicate data for which weight 1 is set, and crosses indicate data for which weight 0 is set.

Focusing on F _motorbike in FIG. 11, the difference between when the weighted discrimination reference value is calculated and when it is not calculated is visualized. FIG. 12 shows the sigmoid function of the label sheep optimized by the discrimination reference value for the label motorbike. Here, the discrimination reference value of F _motorbike > 0 is also a selection target. In FIG. 13, for the same discriminant reference value, based on the first correlation matrix information 52-1 illustrated in FIG. 9, the discriminant reference value of F _motorbike > 0 is not used as the basis for calculating the posterior probability. The part depending on the discriminant reference value of the label motorbike of the sigmoid function of the label sheep In both FIG. 12 and FIG. 13, the discrimination reference value of F _motorbike <0 is selected.

In FIG. 12 and FIG. 13, the circled points are feature amounts (t _sheep = 1) extracted from the image associated with the marker shape, and the crossed points are obtained from the image not associated with the marker shape. The extracted feature amount (t _sheep = 0). The horizontal axis indicates the discrimination reference value F _motorbike corresponding to the learning model of the marker motorbike. In the example of FIG. 13, the weight of the discrimination reference value is 1 for data of F _motorbike <0. For data of F _motorbike > 0, the weight of the discrimination reference value is 0.

Under the influence of FIG. 12, data corresponding to _{F motorbike>} 0 (the right half of the data in FIG. 12), characteristic of sigmoid function becomes a gentle curve than sigmoid function of Fig. _13, the value of _{F motorbike} is _{P sheep} It can be seen that the effect on the value is small. On the other hand, in FIG. 13, it can be seen that the influence of the value F _{motorbike on} the value of P _sheep is greater than that of the sigmoid function of FIG. In addition, as shown in FIG. 11, in this comparison result, the recognition of the marker shape is performed when the weighted discrimination reference value is calculated using the first correlation matrix information 52-1, rather than when it is not performed. The rate is high.

  Next, an example of a comparison result between the case where the weighted determination reference value is calculated using the second correlation matrix information 52-2 and the case where the weighted determination reference value is not calculated will be described. FIG. 14 shows a comparison result of the recognition rate of the sign when the weighted discrimination reference value is calculated using the second correlation matrix information 52-2 and when it is not calculated. In the comparison performed in FIG. 14, an image having a discrimination reference value smaller than 0 was used. In FIG. 14, circles indicate data for which 1 is set as the weight, and crosses indicate data for which 0 is set as the weight.

Focusing on F _person in FIG. 14, the difference between the case where the weighted discrimination reference value is calculated and the case where it is not calculated is visualized. FIG. 15 shows the sigmoid function of the label cat optimized by the discrimination reference value for the label person. Here, a discrimination reference value of _Fperson <0 is also a selection target. In FIG. 16, for the same discriminant reference value, based on the second correlation matrix information 52-2 illustrated in FIG. 15, the discriminant reference value of F _person <0 is not used as the basis for calculating the posterior probability. The part depending on the discriminant reference value of the label person of the sigmoid function of the label cat is shown. In FIG. 15 and FIG. 16, the discrimination reference value of F _person > 0 is selected.

In FIGS. 15 and 16, the circled points are feature quantities (t _cat = 1) extracted from the image associated with the marker cat, and the crossed dots are from the image not associated with the marker cat. The extracted feature amount (t _cat = 0). The horizontal axis indicates the discrimination reference value _Fperson corresponding to the learning model of the label person.

In the example of FIG. 16, the weight of the discrimination reference value is 1 for data with F _person > 0. For data of _Fperson <0, the weight of the discrimination reference value is 0.

In FIG. 15, due to the influence of data corresponding to F _person <0 (the data in the left half of FIG. 15), the characteristics of the sigmoid function become a gentler curve than the sigmoid function of FIG. 16, and the value of F _person is P _cat It can be seen that the influence on the value of is small. On the other hand, FIG. 16 shows that the influence of the value of F _{person on} the value of P _cat is larger than that of the sigmoid function of FIG. Further, as shown in FIG. 14, in this comparison result, the marker cat is recognized when the weighted discrimination reference value is calculated using the second correlation matrix information 52-2 than when it is not calculated. The rate is high.

  FIG. 17 shows a specific example of the parameter tilde A included in the probability model information 54 in this comparison. FIG. 17 also shows that label correlation is utilized in this comparison. For example, since the value of the parameter tilde A in the row bus and the column car is negative and the absolute value thereof is large, it can be seen that the positive correlation between the sign bus and the sign car can be expressed. Similarly, since the value of the parameter tilde A of row bus and column motorbike is positive and the absolute value thereof is also large, it can be seen that the negative correlation between the sign bus and the sign motorbike can be expressed. Thereby, the discrimination performance of the image associated with the sign bus is high. On the other hand, the value of the parameter tilde A in the row bus and the column car is close to zero, which indicates that there is no correlation between the label dog and the label car.

  In addition, this invention is not limited to the above-mentioned embodiment. Moreover, the above specific numerical values and character strings are examples, and are not limited to these numerical values and character strings.

  DESCRIPTION OF SYMBOLS 10 Information processing apparatus, 12 Control part, 14 Storage part, 16 User interface (UI) part, 20 Image reception part, 22 Learning model information storage part, 24 Correlation matrix information storage part, 26 Probability model information storage part, 28 Feature-value Extraction unit, 30 Discrimination reference value calculation unit, 32 Discrimination reference value weighting unit, 34 Dependence possibility calculation unit, 36 Classification determination unit, 38 Output unit, 40 Learning unit, 50 Learning model information, 52 Correlation matrix information, 54 Probability Model information.

Claims (3)

Image receiving means for receiving an image that is a target for specifying at least one of a plurality of classifications;
For each of the plurality of classifications, the image receiving means is based on the feature quantity of the image received by the image receiving means and the feature quantities of a plurality of images belonging to at least one of the plurality of classifications. A discriminant reference value specifying means for specifying a discriminant reference value serving as a reference for discriminating whether or not the received image belongs to the classification;
For each of the plurality of classifications, a plurality of classification criteria values that determine whether or not they belong to the classifications, and a plurality of classifications that belong to at least one of the plurality of classifications that are selected based on the discrimination criterion values First correlation information that is specified based on the image and represents the possibility that an image belonging to the category belongs to another category, or a possibility that an image not belonging to the category belongs to the other category Weighted discrimination reference value specifying means for specifying a weighted discrimination reference value associated with a combination of the classification and the other classification based on any one of the second correlation information to be expressed,
For each of the plurality of classifications, the image received by the image receiving unit is based on the weighted determination reference value associated with the combination of the classification specified for each of the plurality of other classifications and the other classification. Classification categorical possibility determination means for specifying a value indicating the likelihood of belonging to a classification,
An output means for outputting information indicating at least one classification to which an image received by the image receiving means is specified, which is specified based on the value specified by the classification affiliation possibility specifying means;
A program characterized by causing a computer to function. The image that the image receiving means accepts based on the combination of the correlation information and the discrimination reference value for each of the classifications that is different from the classification for each classification, Identify values that are likely to belong to
The program according to claim 1. Image receiving means for receiving an image that is a target for specifying at least one of a plurality of classifications;
For each of the plurality of classifications, the image receiving means is based on the feature quantity of the image received by the image receiving means and the feature quantities of a plurality of images belonging to at least one of the plurality of classifications. A discrimination reference value specifying means for specifying a discrimination reference value serving as a reference for discriminating whether or not the received image belongs to the classification;
For each of the plurality of classifications, a plurality of classification criteria values that determine whether or not they belong to the classifications, and a plurality of classifications that belong to at least one of the plurality of classifications that are selected based on the discrimination criterion values First correlation information that is specified based on the image and represents the possibility that an image belonging to the category belongs to another category, or a possibility that an image not belonging to the category belongs to the other category Weighted discrimination reference value specifying means for specifying a weighted discrimination reference value associated with a combination of the classification and the other classification based on any one of the second correlation information to be expressed;
For each of the plurality of classifications, the image received by the image receiving unit is based on the weighted determination reference value associated with the combination of the classification specified for each of the plurality of other classifications and the other classification. A classification affiliation possibility identification means for identifying a value indicating the level of possibility of belonging to a classification,
Output means for outputting information indicating at least one classification to which an image received by the image receiving means, specified based on the value specified by the classification affiliation possibility specifying means,
An information processing system comprising:

Images