Information processing system, information processing device, information processing method, and storage medium

Abstract

This information processing system includes: an extraction means for extracting a biological characteristic amount from biological information; an extension means for changing the positional relationship, in a characteristic amount space, between the biological characteristic amount and a representative characteristic amount representing the biological characteristic amount for individuals calculated on the basis of correct answer information indicating the individual to which the biological information belongs and the biological characteristic amount extracted from the biological information; and a parameter update means for updating a parameter of a learning model that constitutes the extraction means, such that the distance, in the characteristic amount space, between the biological characteristic amount of a first individual and the representative characteristic amount pertaining to the first individual becomes smaller when the positional relationship is changed, and such that the distance between the biological characteristic amount of the first individual and the representative characteristic amount pertaining to the other individuals becomes larger.

Description

Information processing system, information processing device, information processing method, and storage medium

[0001] The present disclosure relates to an information processing system, an information processing device, an information processing method, and a storage medium.

[0002] There are technologies that authenticate individuals by identifying them from their biometric information. To identify individuals from their biometric information, a learning model constructed by learning how to extract features from each individual's biometric information may be used. In feature extraction training, multiple biometric features and correct labels (correct information indicating which individual each belongs to) are provided as training data. Features classified into the same class (features obtained from the biometric information of the same person) are trained to be closer in feature space, while features classified into different classes are trained to be farther apart. During the training process, a loss function is used to calculate the difference between the result calculated by the learning model during training and the correct answer, and the training progresses to minimize the difference calculated by the loss function. For example, distance learning using the angle between different feature vectors uses a loss function that uses a softmax function such as ArcFace or CosFace. Patent Document 1 discloses a method in which, if the angle θ between a feature vector and a weight vector of a correct class is smaller than a predetermined value, a margin-adding process by ArcFace using the feature vector and the weight vector is applied to calculate the value of a loss function, and if the angle θ is equal to or greater than the predetermined value, a margin-adding process by CosFace using the feature vector and the weight vector is applied to calculate the value of the loss function, and a learning model is trained by a gradient method based on the calculated value of the loss function.

[0003] The loss function value is calculated based on the distance (angle) between the features of an individual's biometric information and the representative features of the person's own class representing that individual, and the distance between the representative features of the other person's class other than that individual. For a given individual's features, there is other person's data with various degrees of similarity. However, when training with a loss function using softmax, the individual data is not directly compared with each other, but only with the representative features of each other. As a result, there is a lack of diversity in the features of the other person's data compared with the features of the individual compared during training, which could result in a training model that does not improve discrimination performance for other classes.

[0004] Japanese Patent Application Laid-Open No. 2022-180119

[0005] One of the aims of this disclosure is to improve upon the techniques disclosed in the prior art documents.

[0006] According to one aspect of the disclosure, an information processing system includes an extraction means for extracting biometric features from biometric information; an extension means for changing a positional relationship in a feature space between a representative feature representing the biometric features of each individual calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information; and a parameter update means for updating parameters of a learning model constituting the extraction means so that when the positional relationship is changed, the distance in the feature space between the biometric feature of a first individual and the representative feature related to the first individual becomes smaller and the distance between the biometric feature of the first individual and the representative feature related to another individual becomes larger.

[0007] According to one aspect of the disclosure, an information processing device includes an extraction means for extracting biometric features from biometric information, an extension means for changing a positional relationship in a feature space between a representative feature representing the biometric features of each individual calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features, and a parameter update means for updating parameters of a learning model constituting the extraction means so that when the positional relationship is changed, the distance in the feature space between the biometric feature of a first individual and the representative feature related to the first individual becomes smaller and the distance between the biometric feature of the first individual and the representative feature related to another individual becomes larger.

[0008] According to one aspect of the disclosure, an information processing method extracts biometric features from biometric information, changes a positional relationship in a feature space between a representative feature that represents the biometric features of each individual, the representative feature being calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features, and updates parameters of a learning model that constitutes the extraction means so that when the positional relationship is changed, the distance in the feature space between the biometric feature of a first individual and the representative feature related to the first individual becomes smaller, and the distance between the biometric feature of the first individual and the representative feature related to another individual becomes larger.

[0009] According to one aspect of the disclosure, a storage medium stores a program that causes a computer to execute a process of extracting biometric features from biometric information, changing a positional relationship in a feature space between a representative feature that represents the biometric features of each individual calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features, and updating parameters of a learning model that constitutes the extraction means so that when the positional relationship is changed, the distance in the feature space between the biometric feature of a first individual and the representative feature related to the first individual becomes smaller and the distance between the biometric feature of the first individual and the representative feature related to another individual becomes larger.

[0010] FIG. 1 is a first diagram showing an example of a learning system. FIG. 2 is a diagram showing an example of a loss function. FIG. 3 is a first diagram showing an example of an iris feature extension means. FIG. 4 is a second diagram showing an example of an iris feature extension means. FIG. 5 is a first diagram showing an example of an extended iris feature. FIG. 6 is a third diagram showing an example of an iris feature extension means. FIG. 7 is a second diagram showing an example of an extended iris feature. FIG. 8 is a diagram showing an example of a representative feature extension means. FIG. 9 is a diagram showing an example of an extended representative feature. FIG. 10 is a second flowchart showing an example of a learning process. FIG. 11 is a second diagram showing an example of an example of a learning system. FIG. 12 is a diagram showing an example of a loss function calculation means. FIG. 13 is a third flowchart showing an example of a learning process. FIG. 14 is a diagram showing an example of an authentication system. FIG. 15 is a flowchart showing an example of an authentication process. FIG. 16 is a first diagram showing a configuration of an information processing system having a minimum configuration. FIG. 17 is a first flowchart showing an example of the operation of an information processing system having a minimum configuration. FIG. 18 is a second diagram showing a configuration of an information processing system having a minimum configuration. FIG. 19 is a second flowchart showing an example of the operation of an information processing system having a minimum configuration. FIG. 20 is a third diagram showing a configuration of an information processing system having a minimum configuration. 10 is a third flowchart illustrating an example of the operation of an information processing system having a minimum configuration.

[0011] First Embodiment A learning system according to the first embodiment will be described below with reference to FIGS. 1 to 3. (Configuration) FIG. 1 is a first diagram showing an example of a learning system 100. The learning system 100 learns a method for extracting features from biometric information that will improve identification accuracy. The following description will use an iris or an image of an iris as an example of biometric information, but the learning system 100 according to each embodiment can also be applied to extracting features from various types of biometric information other than iris, such as face, fingerprint, vein, palm print, ear acoustics, voice, and images containing these.

[0012] The learning system 100 includes a learning data acquisition means 11 , an iris feature extraction means 12 , an iris feature extension means 13 , a loss function calculation means 14 , a gradient calculation means 15 , and a parameter update means 16 .

[0013] The training data acquiring means 11 acquires the training data 10. The training data 10 is, for example, data that combines an image including an iris with correct answer information (label) indicating whose iris is in the image.

[0014] The iris feature extraction means 12 extracts iris feature amounts from an image. The extracted feature amounts are represented as vectors, matrices, or tensors. The iris feature amounts can be used for iris authentication. The iris feature extraction means 12 is configured, for example, by a neural network. The iris feature extraction means 12 extracts iris feature amounts from an input image. The iris feature extraction means 12 is an example of an extraction means.

[0015] The iris feature expansion means 13 expands the iris feature extracted by the iris feature extraction means 12 to generate expanded iris feature vectors. "Expansion" refers to changing the vector of the iris feature extracted from an image of individual A so that the positional relationship between the iris feature extracted from the image of individual A and the representative vector representing the class of individual A generated by learning becomes greater. The positional relationship refers to the distance or angle between data in feature space and represents the similarity between the data. For example, the iris feature expansion means 13 expands the extracted iris feature vector so that the angle between the representative vector of individual A and the vector of the iris feature of individual A extracted by the iris feature extraction means 12 becomes greater. Specific examples of the expansion method will be described later (second to fourth embodiments). By generating expanded iris feature vectors that are farther away from the representative vector and learning that this data also belongs to individual A's class, improved identification accuracy can be expected. For example, if the position of individual A's features is shifted (extended) to be closer to the features of individual B, which is difficult to distinguish from individual A, and the system is trained to recognize the extended individual A's features as belonging to the individual A's class, it is expected that the performance of distinguishing between the individual A and other data, which is difficult to distinguish from the individual A, will be improved. If the iris feature extension means 13 generates extended iris features so that the positional relationship with the representative vector of the individual class is greater, the similarity calculated from the generated extended iris features will be lower with the individual class compared to the similarity that can be calculated from the original iris features, and will vary depending on the relative positional relationship with the representative feature of each class, and can take higher or lower values ​​with other classes. By generating extended iris features in this way, it is possible to diversify the features compared during learning, which is expected to improve learning accuracy. The iris feature extension means 13 is an example of an extension means.

[0016] The loss function calculation means 14 calculates the loss using a loss function from the correct answer information and representative feature corresponding to the extended iris feature. Generally, a loss function is configured to calculate the loss, which is the difference between the value predicted by the learning model and the correct answer, and learning is performed to minimize this loss. In this embodiment, the loss is calculated using a loss function using a softmax function such as ArcFace or CosFace. These loss functions express the distance between the value predicted by the learning model (e.g., the iris feature vector extracted from the image of individual A) and the correct answer (e.g., the representative vector of individual A's class) as an angle. An example of a loss function is shown in FIG. 2. Let θy be the angle between the iris feature vector extracted from the input image and the representative vector of the person's class (a class composed of the person's iris feature), and θc ≠ y be the angle between the iris feature vector extracted from the input image and the representative vector of the other person's class (a class generated for each other person composed of the iris feature of other people). The loss function using ArcFace can be expressed by equation (1) in FIG. 2. In equation (1), s and m are scale and margin parameters, respectively, and these values ​​are appropriately set to improve learning accuracy. For example, increasing the margin m results in learning parameters that increase the positional distance between other classes and reduce intra-class variance. The loss function calculation means 14 calculates the loss, for example, using equation (1) in FIG. 2. The loss function calculation means 14 calculates the loss not only for the iris feature extracted from the image of individual A, but also for the extended iris feature generated by the iris feature extension means 13. In this case, the loss function calculation means 14 calculates the loss by setting θy as the angle between the vector of individual A's extended iris feature and the representative vector of individual A's true class, and setting θc ≠ y as the angle between the vector of individual A's extended iris feature and the representative feature vector of the other class. The true class and other class are generated in the process of training the neural network constituting the iris feature extraction means 12. The feature quantity extracted by the iris feature extraction means 12 is a vector representing a class classification result indicating which class the extracted feature quantity is classified into. At the beginning of learning, a random value is extracted from the vector representing the class classification result, but as learning progresses, a value close to the correct one-hot vector is output.

[0017] The gradient calculation means 15 uses the loss calculated by the loss function calculation means 14 to calculate the gradient of the loss for each parameter of the neural network that constitutes the iris feature extraction means 12 .

[0018] The parameter update means 16 updates the parameters based on the gradient calculated by the gradient calculation means 15 and a preset learning rate. The parameter update means 16 updates each parameter of the neural network that constitutes the iris feature extraction means 12 so as to reduce the loss calculated by the loss function calculation means 14.

[0019] (Operation) FIG. 3 shows an example of the learning process performed by the learning system 100. The learning data acquisition means 11 acquires a learning dataset (step S1). The learning data acquisition means 11 holds the acquired large number of learning data 10. Next, the learning data acquisition means 11 acquires a learning data batch (step S2). For example, the learning data acquisition means 11 randomly acquires a mini-batch of learning data 10 from the dataset acquired in step S1. The learning data acquisition means 11 acquires a predetermined number of iris images and their labels (correct answer information), defined as a batch size. The batch size may be any positive integer. It is also assumed that the individuals to be trained are set in advance. Next, the iris feature extraction means 12 extracts iris features (step S3). The iris feature extraction means 12 extracts iris features from each image included in the learning data batch acquired in step S2. Next, the iris feature extension means 13 generates extended iris features (step S4). The iris feature expansion means 13 generates extended iris features from the iris features extracted in step S3. The iris feature expansion means 13 generates extended iris features for iris features related to the individual to be trained, among the iris features extracted in step S3. The iris feature expansion means 13 may generate extended iris features for all of the iris features related to the individual to be trained, or may generate extended iris features from only a portion of them. Next, the loss function calculation means 14 calculates a loss using a loss function for the iris features extracted in step S3 and the extended iris features generated in step S4 (step S5). Next, the gradient calculation means 15 calculates the gradient of the loss function by, for example, differentiating the value of the loss function with the parameters of the neural network (step S6). Next, the parameter update means 16 updates the parameters of the neural network using the gradient and a predetermined learning rate (step S7).

[0020] The learning system 100 optimizes the parameters by repeatedly executing the processes of steps S2 to S7 (referred to as a learning loop). For example, the learning system 100 terminates the learning loop after repeating the learning loop a predetermined number of times. Alternatively, the learning system 100 may terminate the learning loop when the loss value calculated in step S4 no longer changes. Because the extended iris feature values ​​generated in the learning loop are different each time, the relative positional relationship between the extended iris feature values ​​and the representative feature values ​​of each class is different each time. This makes it possible to learn various positional relationships with other classes, improving the performance of distinguishing between other classes. After completing the learning loop processes of steps S2 to S7, the iris feature extraction means 12 saves the parameters (step S8).

[0021] (Effects) According to the first embodiment, extended features that are close to the biometric features of other people or that are distant from the features that represent the biometric features of the person are generated from biometric features extracted from the person's biometric information, and the system learns that the extended features belong to the person's class. This makes it possible to build a learning model that has improved performance in distinguishing between data of other people that is difficult to distinguish from the person, compared to when extended features are not generated.

[0022] Second Embodiment A learning system 100 according to the second embodiment will be described with reference to FIG. 4. (Configuration) The learning system 100 according to the second embodiment has an iris feature expansion means 13a instead of the iris feature expansion means 13 of the first embodiment. The other configurations are the same as those of the first embodiment. An example of the iris feature expansion means 13a is shown in FIG. 4. The iris feature expansion means 13a has a random number generation means 131a and an iris feature conversion means 132a. The random number generation means 131a generates random numbers. The iris feature conversion means 132a uses the random numbers generated by the random number generation means 131a to correct the iris feature amounts extracted by the iris feature extraction means 12 and generate expanded iris feature amounts. The iris feature expansion means 13a is an example of an expansion means.

[0023] Specifically, the random number generating means 131a generates, for example, random numbers with the same number of dimensions as the iris feature amount, and generates a feature amount r which is a vector or matrix having the generated random numbers as elements. If the iris feature amount extracted by the iris feature extracting means 12 in step S3 is x, the iris feature converting means 132a generates an extended iris feature amount x' using a predetermined function f according to the following equation (2): x' = f (x, r) ... (2) For example, the iris feature converting means 132a may generate the extended iris feature amount x' according to the following equation (2'): x' = (x + r) / (|x + r|) ... (2') However, the feature amount r generated by the random number may also be converted according to the following equation (2") so that the norm of the feature amount r generated by the random number becomes a value a which is sufficiently smaller than the iris feature amount: r' = r a / |r| ... (2'')

[0024] (Operation) The learning process of the parameters of the iris feature extraction means 12 by the learning system 100 according to the second embodiment is the same as the process described with reference to Fig. 3. In the second embodiment, in step S4, the extended iris feature amount is generated using the above formula (2) or the like.

[0025] (Effects) According to the second embodiment, the same effects as those of the first embodiment can be obtained. Furthermore, according to the second embodiment, the extended characteristic feature can be generated by the above formula (2) or the like. Note that the above formula is an example. The extended characteristic feature may be generated by another formula. Furthermore, an embodiment may be adopted in which the feature r is generated using one or more predetermined values ​​without generating random numbers.

[0026] Third Embodiment A learning system 100 according to the third embodiment will be described with reference to FIGS. 5 and 6. (Configuration) The learning system 100 according to the third embodiment includes an iris feature expansion unit 13b instead of the iris feature expansion unit 13 of the first embodiment. Other configurations are the same as those of the first embodiment. An example of the iris feature expansion unit 13b is shown in FIG. 5. The iris feature expansion unit 13b includes a representative feature storage unit 131b and an iris feature conversion unit 132b. The representative feature storage unit 131b stores representative features (representative vectors) for each class. The representative feature storage unit 131b receives correct answer information (personal class) corresponding to the iris feature as input, selects representative features for the corresponding class (representative vectors for the personal class), and outputs the selected representative features to the iris feature conversion unit 132b. The iris feature conversion unit 132b converts the iris feature extracted by the iris feature extraction unit 12 using the representative features for the personal class to generate extended iris features. For example, the iris feature conversion means 132b is a function that generates an extended iris feature from the representative feature of the person class and the iris feature. The iris feature extension means 13b is an example of an extension means.

[0027] Specifically, the iris feature conversion means 132b calculates a plane spanned by the vector of the extracted iris feature and the representative vector of the true person class. The calculated plane 60 is shown in FIG. 6 . Vector 61 is the representative vector of the true person class, and vector 62 is the extracted iris feature. The iris feature conversion means 132b generates a vector as an extended iris feature 63 by rotating vector 62 on plane 60 by an angle m in a direction away from representative vector 61 of the true person class. The angle m may be a predetermined fixed value, or may be calculated using the following formula (3) using the variance σ of the similarity between the true person representative feature and multiple samples (iris features belonging to the true person class) and a constant α: m = α / σ (3) By calculating the angle m representing the margin using formula (3), the angle m is small in the early stages of learning when the variance σ of the true person samples is large, and the margin becomes larger in the later stages of learning when the variance σ becomes smaller. Changing the margin in this way as the learning progresses enables stable learning.

[0028] (Operation) The learning process of the parameters of the iris feature extraction means 12 by the learning system 100 according to the third embodiment is the same as the process described with reference to Fig. 3. In the third embodiment, in step S4, the iris feature extension means 13b generates extended iris features by rotating the vector direction of the extracted iris features by a predetermined angle in a direction away from the vector direction of the person's representative feature.

[0029] (Effects) According to the third embodiment, the same effects as those of the first embodiment can be obtained. Specifically, according to the third embodiment, the representative vector of the true person class is used as a reference and extended features that deviate from this reference are generated, which makes it possible to diversify the features to be learned and improve the performance of distinguishing the true person from features of others that are difficult to distinguish from the true person.

[0030] Fourth Embodiment A learning system 100 according to the fourth embodiment will be described with reference to FIGS. 7 and 8. (Configuration) The learning system 100 according to the fourth embodiment includes an iris feature expansion unit 13c instead of the iris feature expansion unit 13 of the first embodiment. Other configurations are the same as those of the first embodiment. An example of the iris feature expansion unit 13c is shown in FIG. 7. The iris feature expansion unit 13c includes a representative feature storage unit 131c and an iris feature conversion unit 132c. The representative feature storage unit 131c stores representative features (representative vectors) of each class. The representative feature storage unit 131c receives correct answer information (the subject's own class) corresponding to the iris feature as input, selects representative features (representative vectors of other classes) of classes other than the class corresponding to the correct answer information, and outputs them to the iris feature conversion unit 132c. In the case of a child, the representative feature storage unit 131c may arbitrarily select a representative feature from the representative features of other classes. Alternatively, the representative feature storage means 131c may select the representative feature of the class that is most difficult to distinguish from the representative feature of the true class from all classes (for example, the class with the closest distance or angle), or may arbitrarily select a representative feature from the top K classes that are difficult to distinguish. The iris feature conversion means 132c converts the iris feature extracted by the iris feature extraction means 12 using the representative feature of the other class to generate an extended iris feature. For example, the iris feature conversion means 132c is a function that generates an extended iris feature from the representative feature of the other class and the iris feature. The iris feature expansion means 13c is an example of an expansion means.

[0031] Specifically, the iris feature conversion means 132c calculates a plane spanned by the vector of the extracted iris feature amount and the representative vector of the other class. The calculated plane 80 is shown in Figure 8. Vector 81 is the representative vector of the true class, vector 82 is the extracted iris feature amount, and vector 83 is the representative vector of the selected other class. The iris feature conversion means 132c generates a vector obtained by rotating the direction of vector 82 on plane 80 by angle m in a direction approaching representative vector 83 of the other class as extended iris feature amount 84. The angle m may be a predetermined fixed value, or may be calculated using the above equation (3).

[0032] (Operation) The learning process of the parameters of the iris feature extraction means 12 by the learning system 100 according to the fourth embodiment is similar to the process described with reference to Fig. 3. In the fourth embodiment, in step S4, the iris feature extension means 13c generates an extended iris feature by rotating the iris feature vector by a predetermined angle so that the vector direction of the extracted iris feature approaches the direction of the representative vector of another person.

[0033] (Effects) According to the fourth embodiment, the same effects as those of the first embodiment can be obtained. Specifically, according to the fourth embodiment, the representative feature vector of another person is used as a reference and extended features are generated to approximate the reference, which makes it possible to diversify the features to be learned and improve the performance of distinguishing between the features of another person who is difficult to distinguish from the person himself / herself.

[0034] Fifth Embodiment A learning system 100 according to the fifth embodiment will be described with reference to FIGS. 9 to 11. (Configuration) The learning system 100 according to the fifth embodiment has a representative feature extension means 13d instead of the iris feature extension means 13 of the first embodiment, and a loss function calculation means 14d instead of the loss function calculation means 14. The other configurations are the same as those of the first embodiment. An example of the representative feature extension means 13d and the loss function calculation means 14d is shown in FIG. 9. The representative feature extension means 13d has a representative feature storage means 131d, a representative feature extension means 132d, and a loss weight calculation means 133d. The representative feature storage means 131c stores the representative feature amount of each class.

[0035] The representative feature extension means 132d receives as input the representative feature of each class acquired from the representative feature storage means 131c and correct answer information (true class) corresponding to the iris feature, and generates extended representative features. Specifically, the representative feature extension means 132d calculates the distance and similarity between the true class and the other class to be learned. Then, based on the correct answer information, the representative feature extension means 132d selects the representative feature of the true class and the representative feature of the other class that is closest to the representative feature of the true class (small distance, high similarity). The representative feature extension means 132d calculates a plane spanned by the representative vector of the true class and the representative vector of the other class that is closest to the representative vector of the true class. The calculated plane 100A is shown in FIG. 10. Vector 101 is the representative vector of the true class, vector 102 is the extracted iris feature, and vector 103 is the representative vector of the other class that is closest to the representative vector of the true class. The representative feature extension means 132d generates a vector 104 indicating an extended representative feature by rotating the representative vector 101 of the true class by an angle m so as to bring it closer to the representative vector 103 of the other class. The vector 104 is closer to the representative vector 103 of the other class than the representative vector 101 of the true class. Therefore, the position of the representative feature is learned so as to separate these two classes in particular, improving the recognition performance for hard negative samples. The representative feature extension means 13d is an example of an extension means.

[0036] The loss weight calculation means 133d receives as input the extended representative features generated by the representative feature extension means 132d and the representative features of each class, and calculates a loss weight. (Step 1) First, the loss weight calculation means 133d calculates the distance (such as cosine similarity) between, for example, the extended representative features of a certain individual A and the representative features of each class. In this case, for individual A, the distance between the representative features of individual A before extension and the extended representative features of individual A is calculated, and for other individuals, the distance between the extended representative features of individual A and the representative features of each individual is calculated. (Step 2) Next, the loss weight calculation means 133d calculates a value normalized by a softmax function for the calculated distances to the representative features of each class. When each calculated distance (the distance between the extended representative features of individual A and the representative features of the self and other classes) is input to the softmax function, each distance is converted to a number between 0 and 1, and the sum of the converted values ​​is 1. (Step 3) Finally, the loss weight calculation means 133d calculates the value q after conversion by the softmax function for the self class. ｙ is extracted and used as the loss weight. The larger the margin m (angle m) when calculating the extended representative feature, the smaller the loss weight becomes. This reduces the impact even if the extended representative feature vector 104 deviates significantly from the original person class representative vector 101, making it possible to achieve stable convergence of the loss function.

[0037] The loss function calculation means 14d calculates the extended representative feature and the loss weight q ｙ and the iris feature amount extracted from the image are received as input, and the loss is calculated using the loss function of, for example, the following equation (4).

[0038]

[0039] Comparing with equation (1) in FIG. 2, the sum of the angle θy between the vector of the iris feature amount and the representative vector of the person class and the margin m is replaced with the angle θy' between the vector 102 of the iris feature amount and the vector 104 of the extended representative feature amount, and the loss weight q ｙ is multiplied as a coefficient.

[0040] (Operation) Figure 11 shows an example of a learning process by the learning system 100 according to the fifth embodiment. The same processes as in Figure 3 are assigned the same reference numerals and will be briefly described. The learning data acquisition means 11 acquires a learning dataset (step S1). Next, the learning data acquisition means 11 acquires a learning data batch (step S2). Next, the iris feature extraction means 12 extracts iris features (step S3). The iris feature extraction means 12 extracts iris features from each image included in the learning data batch acquired in step S2. Next, the representative feature extension means 132d generates extended representative features (step S4d1). The representative feature extension means 132d rotates the representative vector of the true class by a predetermined angle so that the direction of the representative vector of the true class of the learning target approaches the direction of the representative vector of the other class, thereby generating extended representative features whose distance from the true class's iris features has been changed. Next, the loss weight calculation means 133d calculates the loss weight q ｙ (Step S4d2). The loss weight calculation means 133d calculates the distance between the extended representative feature and all classes, and converts the distance between all classes using a softmax function (q １ , q ２ , ..., q Ｃ ), the value q representing the distance from the representative vector of the person's class ｙ Next, the loss function calculation means 14d extracts the iris feature extracted in step S3, the extended representative feature generated in step S4d1, and the loss weight q ｙ (Step S5d) calculates the loss using the loss function of the above equation (4). Next, the gradient calculation means 15 calculates the gradient of the loss function (Step S6). Next, the parameter update means 16 updates the parameters of the iris feature extraction means 12 (Step S7). The learning system 100 repeats the learning loop of Steps S2 to S7 to optimize the parameters. When the learning loop process ends, the iris feature extraction means 12 saves the parameters (Step S8).

[0041] (Effects) According to the fifth embodiment, an extended representative feature is generated by approximating the representative feature of the true person class to the representative feature of the other person class, and by learning that the iris feature extracted from the image belongs to the true person class whose representative vector is the extended representative feature, parameters are learned that separate the extended representative feature from the representative feature of the other person class. This makes it possible to construct a learning model that improves the performance of distinguishing between the true person and other person data that is difficult to distinguish from the true person. The fifth embodiment can be combined with any of the first to fourth embodiments. In addition, the loss weight q ｙ By calculating the loss using the formula (1), it is possible to realize stable convergence of the loss function. In the fifth embodiment, the loss function calculation means 14 of the first embodiment or the like may be provided instead of the loss function calculation means 14d.

[0042] Sixth Embodiment A learning system 100e according to a sixth embodiment will be described with reference to Figs. 12 to 14. (Configuration) Fig. 12 is a diagram showing an example of the learning system 100e. The learning system 100e includes a learning data acquisition means 11, an iris feature extraction means 12, a loss function calculation means 14e, a gradient calculation means 15, and a parameter update means 16.

[0043] The learning data acquisition means 11, iris feature extraction means 12, gradient calculation means 15, and parameter update means 16 are the same as those in the first embodiment, and therefore description thereof will be omitted. Also, while Fig. 12 illustrates a configuration that does not include the iris feature extension means 13 to 13c according to the first to fourth embodiments or the representative feature extension means 13d according to the fifth embodiment, a configuration that includes the iris feature extension means 13, etc., may also be used. In other words, the sixth embodiment can be combined with the first to fifth embodiments.

[0044] An example of the loss function calculation means 14e is shown in FIG. 13. The loss function calculation means 14e includes a similarity calculation means 141e, a margin calculation means 142e, and a loss calculation means 143e. The similarity calculation means 141e receives iris feature values ​​as input and calculates the similarity (distance and angle) between the iris feature values ​​and the representative vectors of each class. The similarity calculation means 141e outputs the calculated similarity to the margin calculation means 142e. The margin calculation means 142e calculates a margin to be assigned to each class based on the similarity with each class and the correct answer information. The margin calculation means 142e calculates a margin for each class according to the similarity. By assigning the calculated margin to each class, the classification performance for classes that are difficult to classify is improved. For example, if a large margin is assigned to a certain class, the classification performance for that class also improves significantly. The classification performance can be improved by setting a small margin for other classes that are already easy to classify (other classes with low similarity to begin with) and by setting a large margin for other classes that are difficult to classify (other classes with high similarity). Other classes with high similarity are classes that are prone to misclassification. Therefore, in order to particularly improve the classification performance for other classes with high similarity, a larger margin may be set for higher similarity. For example, the margin m may be determined using the following linear equation (5) for the normalized similarity x (0 to 1): m = ax + b (5) Equation (5) for calculating the margin m is an example. As another example, m may be calculated using a higher-order power function or exponential function.

[0045] The loss calculation means 143e calculates the loss using a loss function, taking into account the margin for each class calculated by the margin calculation means 142e (added to the similarity of the other class). The margin for the person may be fixed. For example, assume that a similarity Sab (the smaller this value, the higher the discrimination performance) is obtained between a feature fa extracted from an image belonging to the class of person A and a representative feature fb belonging to the class of person B. In this case, if the margin m calculated by the margin calculation means 142e for these two classes is mab, the loss calculation means 143e calculates the loss using Sab + mab as the similarity. By adding the margin mab to the similarity between the feature of the class of person A and the feature of the class of person B, the loss is emphasized, and it is expected that the discrimination between person A and person B will be improved. An example of the loss function used to calculate the loss is shown in the following equation (6).

[0046]

[0047] As in equation (1) in FIG. 2, the angle θy in equation (6) is the angle between the iris feature amount and the representative vector of the true class, the angle θc≠y is the angle between the iris feature amount and the representative vector of the other class, s is a scale parameter, and m is a margin parameter of the true class. ｃ is a margin parameter of a certain other class c, and m calculated by the margin calculation means 142e is m in Equation (6). ｃ The margin for other classes is set to m ｃ When m is large, the contribution of the other class to the loss increases. Therefore, when the parameters are updated to reduce the loss, the learning is performed so that the classification with the other class c is improved preferentially. In other words, for the other class c that is likely to be misclassified, m ｃ By setting a large value for , it is possible to improve the discrimination performance from other classes c.

[0048] (Operation) Figure 14 shows an example of the learning process by the learning system 100e. The same processes as in Figure 3 are denoted by the same reference numerals and will be briefly described. The learning data acquisition means 11 acquires a learning dataset (step S1). Next, the learning data acquisition means 11 acquires a learning data batch (step S2). Next, the iris feature extraction means 12 extracts iris features (step S3). Next, the margin calculation means 142e calculates a margin to be assigned to each class based on the similarity between the extracted iris features and each class and the correct answer information (step S4e). Next, the loss calculation means 143e calculates a loss using a loss function (e.g., Equation (6)) using the iris features extracted in step S3, the correct answer information, and the margin for each class calculated in step S4e (step S5e). Next, the gradient calculation means 15 calculates the gradient of the loss function (step S6). Next, the parameter update means 16 updates the parameters of the iris feature extraction means 12 (step S7). The learning system 100 repeats the learning loop of steps S2 to S7 to optimize the parameters. When the learning loop process ends, the iris feature extraction means 12 saves the parameters (step S8).

[0049] (Effects) According to the sixth embodiment, by performing learning by setting a larger margin as the difficulty level of classification increases, it is possible to improve performance for classes that are difficult to classify.

[0050] Seventh Embodiment An authentication system 200 according to the seventh embodiment will be described with reference to Figures 15 and 16. (Configuration) Authentication system 200 includes learning system 100, iris feature extraction means 12' trained by learning system 100, and authentication means 20. The configuration example in Figure 15 describes an example in which authentication system 200 includes learning system 100, but authentication system 200 may include learning system 100e instead of learning system 100.

[0051] The iris feature extraction means 12′ receives as input an iris image of the iris of the person to be authenticated, extracts iris features from the iris image, calculates the distance between the extracted iris features and each class, and outputs identification information of the class with the closest distance to the authentication means 20. This identification information indicates the individual who has the iris features. The authentication means 20 receives the identification information output by the iris feature extraction means 12′ and the identification information of the person to be authenticated, and determines whether they match. The identification information of the person to be authenticated is, for example, the name or identification number of the person to be authenticated obtained based on an account entered by the person to be authenticated or an IC card held by the person to be authenticated. The authentication means 20 determines that the authentication is successful if the identification information output by the iris feature extraction means 12′ matches the identification information of the person to be authenticated obtained from an IC card or the like, and determines that the authentication is unsuccessful if they do not match. The authentication means 20 outputs the authentication result.

[0052] (Operation) FIG. 16 shows an example of the operation of the authentication system 200. An image showing the iris of the person to be authenticated is input to the iris feature extraction means 12′ (step S11). The iris feature extraction means 12′ extracts iris features from the input image (step S12). Next, the iris feature extraction means 12′ identifies the individual indicated by the iris feature and outputs identification information for the identified individual (step S13). The iris feature extraction means 12′ calculates the distance between the extracted iris feature and the features of one or more pre-registered individuals and identifies the class to which the iris feature belongs. The iris feature extraction means 12′ outputs identification information for the individual indicated by the identified class. The authentication means 20 receives the identification information based on the image output by the iris feature extraction means 12′ and performs authentication (step S14). The authentication means 20 outputs the authentication result (step S15). For example, the authentication means 20 outputs whether or not the authentication was successful to another system that uses the authentication result.

[0053] (Effects) According to the seventh embodiment, accurate iris authentication can be achieved by the iris feature extraction means 12' constructed using the learning system 100 of the first to sixth embodiments, etc. For example, even if the authentication subjects include many subjects with similar iris features, these subjects can be identified with high accuracy.

[0054] The learning systems 100 and 100e of the first to sixth embodiments and the authentication system 200 of the seventh embodiment can also be applied to learning systems and authentication systems for identifying not only irises but also other biometric information such as faces, fingerprints, veins, palm prints, ear sounds, and voices.

[0055] 17 is a first diagram showing the configuration of an information processing system having a minimum configuration. The information processing system 800 includes an extraction unit 801, an extension unit 802, and a parameter update unit 803. The extraction unit 801 extracts biometric features from biometric information. The extension unit 802 changes the positional relationship in feature space between a representative feature representing the biometric features of each individual, which is calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features. The parameter update unit 803 updates parameters of a learning model constituting the extraction unit so that the distance in feature space between the representative feature and the biometric features of the first individual is reduced when the positional relationship is changed.

[0056] 18 is a first flowchart showing an example of the operation of an information processing system having a minimum configuration. An extraction unit 801 extracts biometric features from biometric information (step S801). An extension unit 802 generates a representative feature by changing the positional relationship in a feature space between the biometric features and a representative feature that represents the biometric features of each individual, the representative feature being calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information (step S802). The extension unit 802 updates parameters of a learning model constituting the extraction unit so that the distance in the feature space between the representative feature and the biometric feature of the first individual is reduced when the positional relationship is changed (step S803).

[0057] FIG. 19 is a second diagram showing the configuration of an information processing system with a minimum configuration. The information processing system 810 includes an extraction unit 811, a loss function calculation unit 812, and a parameter update unit 813. The extraction unit 811 extracts biometric features from biometric information and generates a representative feature representing a class of the biometric features for each individual. The loss function calculation unit 812 calculates a loss indicating the distance between the biometric features extracted by the extraction unit and the representative feature associated with the biometric features. The parameter update unit 813 updates parameters of a learning model constituting the extraction unit so as to reduce the loss. The loss function calculation unit 812 calculates the loss based on the distance between the representative feature and the biometric feature of a certain individual and the distance between the biometric feature of the certain individual and the representative feature of another individual, setting a larger margin as the distance between the biometric feature of the certain individual and the representative feature of the other individual becomes smaller (the more difficult it is to identify a different individual).

[0058] 20 is a second flowchart illustrating an example of the operation of an information processing system having a minimum configuration. The extraction unit 811 extracts biometric features from biometric information and generates a representative feature representing a class of the biometric features for each individual (step S811). The loss function calculation unit 812 calculates a loss indicating the distance between the biometric features extracted by the extraction unit and the representative feature associated with the biometric features (step S812). The parameter update unit 813 updates the parameters of the learning model constituting the extraction unit so as to reduce the loss (step S813). In step S812, the loss function calculation unit 812 calculates the loss based on the distance between the representative feature and the biometric feature of a certain individual and the distance between the biometric feature of the certain individual and the representative feature of another individual, setting a larger margin as the distance between the biometric feature of the certain individual and the representative feature of another individual decreases.

[0059] 21 is a third diagram showing the configuration of an information processing system having a minimum configuration. The information processing system 820 includes an extraction unit 821, an extension unit 822, and a loss function calculation unit 823. The extraction unit 821 extracts biometric features from biometric information. The extension unit 822 changes the positional relationship in feature space between a representative feature representing the biometric features of each individual, which is calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features. The loss function calculation unit 823 calculates a loss function for training a model that extracts the biometric features based on the distance in the feature space between the representative feature and the biometric features of a first individual when the positional relationship is changed, and the distance between the biometric feature of the first individual and the representative feature of another individual.

[0060] 22 is a third flowchart showing an example of operation of an information processing system having a minimum configuration. The extraction unit 821 extracts biometric features from biometric information (step S821). The extension unit 822 generates a representative feature by changing a positional relationship in a feature space between the biometric features and a representative feature that represents the biometric features of each individual, the representative feature being calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information (step S822). The loss function calculation unit 823 calculates a loss function for training a model that extracts the biometric features, based on a distance in the feature space between the representative feature and the biometric feature of a first individual when the positional relationship is changed, and the distance between the biometric feature of the first individual and the representative feature of another individual (step S823).

[0061] Note that a part of the learning system 100 and the information processing systems 800 and 810 in the above-described embodiments may be realized by a computer. In this case, a program for realizing this function may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be read into the computer system and executed. Note that the "computer system" here refers to the computer system built into the learning system 100 and the information processing systems 800 and 810, and includes hardware such as an OS (Operating System) and peripheral devices.

[0062] Furthermore, "computer-readable recording media" refers to portable media such as flexible disks, optical magnetic disks, ROMs, and CD-ROMs, as well as storage devices such as hard disks built into computer systems. Furthermore, "computer-readable recording media" may also include devices that dynamically store programs for a short period of time, such as communication lines used when transmitting programs over networks like the Internet or communication lines like telephone lines, or devices that store programs for a fixed period of time, such as volatile memory within computer systems that serve as servers or clients in such cases. Furthermore, the above-mentioned programs may be programs that realize some of the aforementioned functions, or may be programs that can realize the aforementioned functions in combination with programs already stored in the computer system.

[0063] Furthermore, part or all of the learning system 100 and information processing systems 800 and 810 in the above-described embodiments may be realized as an integrated circuit such as an LSI (Large Scale Integration). Each functional unit of the learning system 100 and information processing systems 800 and 810 may be individually implemented as a processor, or part or all of them may be integrated into a processor. Furthermore, the integrated circuit method is not limited to LSI, and may be implemented using a dedicated circuit or a general-purpose processor. Furthermore, if an integrated circuit technology that can replace LSI emerges due to advances in semiconductor technology, an integrated circuit based on that technology may be used.

[0064] As described above, several embodiments according to this disclosure have been described, but all of these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope of the invention and its equivalents as defined in the claims, as well as in the scope and spirit of the invention.

[0065] The present disclosure aims to improve the performance of distinguishing between features of a person who is difficult to distinguish from another person, and to provide an information processing system, an information processing method, and a storage medium that can solve the above-mentioned problems.

[0066] <Supplementary Note> The information processing system, information processing device, information processing method, storage medium, and program described in the embodiments can be understood, for example, as follows: Note that the description of the supplementary note corresponding to the dependent claim regarding the information processing system can also be made dependent on the information processing device, information processing method, storage medium, and program.

[0067] (1) An information processing system according to a first aspect includes an extraction means for extracting biometric features from biometric information; an extension means for changing a positional relationship in a feature space between a representative feature representing the biometric features of each individual calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information; and a parameter update means for updating parameters of a learning model constituting the extraction means so that when the positional relationship is changed, the distance in the feature space between the biometric feature of a first individual and the representative feature related to the first individual becomes smaller and the distance between the biometric feature of the first individual and the representative feature related to another individual becomes larger.

[0068] (2) An information processing system according to a second aspect is the information processing system described in (1), further comprising: a loss function calculation means for calculating a loss indicating the difference in distance between the representative feature and the biometric feature of the first individual; and a gradient calculation means for calculating a gradient of the loss with respect to the parameters, and the parameter update means updates the parameters based on the gradient so as to minimize the difference in distance.

[0069] (3) An information processing system according to a third aspect is an information processing system according to any one of (1) to (2), and includes: a means for identifying the individual indicated by the biometric features based on the biometric features extracted by the extraction means and the representative features for each individual; and a means for authenticating the individual based on whether the identified individual matches the individual to be authenticated.

[0070] (4) An information processing system according to a fourth aspect is an information processing system according to any one of (1) to (3), wherein the extension means changes the biometric feature associated with a certain individual to generate an extended biometric feature with a changed distance from the representative feature associated with the individual.

[0071] (5) An information processing system according to a fifth aspect is an information processing system according to any one of (1) to (4), wherein the extension means corrects the biometric feature associated with the individual with a value generated from a random number, thereby generating the extended biometric feature in which the distance from the representative feature associated with the individual is varied in various patterns.

[0072] (6) An information processing system according to a sixth aspect is an information processing system according to any one of (1) to (4), wherein the extension means generates the extended biometric feature by rotating the vector indicated by the biometric feature relating to the individual by a predetermined angle in a direction away from the vector indicated by the representative feature relating to the vector, thereby changing the extended biometric feature so that the distance from the representative feature increases.

[0073] (7) An information processing system according to a seventh aspect is an information processing system according to any one of (1) to (4), wherein the extension means generates the extended biometric feature by rotating the vector indicated by the biometric feature relating to the individual by a predetermined angle in a direction that brings the vector closer to the vector indicated by the representative feature relating to another individual, thereby changing the distance from the representative feature relating to the individual to increase.

[0074] (8) An information processing system according to an eighth aspect is an information processing system according to any one of (1) to (7), wherein the extension means changes the representative feature associated with a certain individual to generate the extended representative feature in which the distance from the biometric feature associated with the individual is changed.

[0075] (9) An information processing system according to a ninth aspect is the information processing system described in (1) to (8), wherein the extension means generates the extended representative feature with the distance changed by rotating the vector indicated by the representative feature related to the individual by a predetermined angle in a direction that brings the vector indicated by the representative feature related to another individual closer to the vector indicated by the representative feature related to another individual.

[0076] (10) An information processing system according to a tenth aspect is the information processing system according to any one of (1) to (9), in which the distance between the extended representative feature and the representative feature for all the individuals is calculated, the calculated distance is transformed using a softmax function, a value for the individual is set as a loss weight from among the distances, and the loss is calculated by multiplying a predetermined loss function by the set loss weight.

[0077] (11) An information processing system according to an eleventh aspect is the information processing system described in (3) to (10) when (2) and (2) are cited, wherein the loss function calculation means calculates the loss based on the distance between the representative feature and the biometric feature of a certain individual and the distance between the biometric feature of the certain individual and the representative feature of another individual, by setting a larger value as a margin as the distance between the biometric feature of the certain individual and the representative feature of another individual is smaller.

[0078] (12) An information processing system according to a twelfth aspect includes an extraction means that extracts biometric features from biometric information and generates a representative feature that represents a class of the biometric features for each individual; a loss function calculation means that calculates a loss indicating a distance between the biometric features extracted by the extraction means and the representative feature related to the biometric features; and a parameter update means that updates parameters of a learning model that constitutes the extraction means so as to reduce the loss, wherein the loss function calculation means calculates the loss based on the distance between the representative feature and the biometric feature of a certain individual and the distance between the biometric feature of the certain individual and the representative feature of another individual, by setting a larger value as a margin as the distance between the biometric feature of the certain individual and the representative feature of the other individual is smaller.

[0079] (13) An information processing system according to a thirteenth aspect includes an extraction means for extracting biometric features from biometric information; an extension means for changing a positional relationship in a feature space between a representative feature representing the biometric features of each individual calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features; and a loss function calculation means for calculating a loss function for training a model for extracting the biometric features, based on a distance in the feature space between the representative feature and the biometric features of a first individual when the positional relationship is changed, and the distance between the biometric feature of the first individual and the representative feature of another individual.

[0080] (14) An information processing device according to a fourteenth aspect includes an extraction means for extracting biometric features from biometric information; an extension means for changing a positional relationship in a feature space between a representative feature representing the biometric features of each individual calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features; and a parameter update means for updating parameters of a learning model constituting the extraction means so that when the positional relationship is changed, the distance in the feature space between the biometric feature of a first individual and the representative feature of the first individual becomes smaller and the distance between the biometric feature of the first individual and the representative feature of another individual becomes larger.

[0081] (15) An information processing device according to a fifteenth aspect includes an extraction means for extracting biometric features from biometric information and generating a representative feature representing a class of the biometric features for each individual; a loss function calculation means for calculating a loss indicating a distance between the biometric features extracted by the extraction means and the representative feature related to the biometric features; and a parameter update means for updating parameters of a learning model constituting the extraction means so as to reduce the loss, wherein the loss function calculation means calculates the loss based on the distance between the representative feature and the biometric feature of a certain individual and the distance between the biometric feature of the certain individual and the representative feature of another individual, by setting a larger value as a margin as the distance between the biometric feature of the certain individual and the representative feature of another individual is smaller.

[0082] (16) An information processing device according to a sixteenth aspect includes an extraction means for extracting biometric features from biometric information; an extension means for changing a positional relationship in a feature space between a representative feature representing the biometric features of each individual calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features; and a loss function calculation means for calculating a loss function for training a model for extracting the biometric features, based on a distance in the feature space between the representative feature and the biometric features of a first individual when the positional relationship is changed, and the distance between the biometric feature of the first individual and the representative feature of another individual.

[0083] (17) An information processing method according to a seventeenth aspect extracts biometric features from biometric information, changes a positional relationship in a feature space between a representative feature that represents the biometric features of each individual, the representative feature being calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features, and updates parameters of a learning model used to extract the biometric features so that when the positional relationship is changed, the distance in the feature space between the biometric feature of a first individual and the representative feature of the first individual becomes smaller and the distance between the biometric feature of the first individual and the representative feature of another individual becomes larger.

[0084] (18) An information processing method according to an eighteenth aspect extracts biometric features from biometric information and generates a representative feature representing a class of the biometric features for each individual, calculates a loss indicating a distance between the biometric features extracted by the extraction means and the representative feature associated with the biometric features, and updates parameters of a learning model to reduce the loss. In calculating the loss, the loss is calculated based on the distance between the representative feature and the biometric feature of a certain individual and the distance between the biometric feature of the certain individual and the representative feature of another individual, with a larger margin set as the distance between the biometric feature of the certain individual and the representative feature of another individual is smaller.

[0085] (19) An information processing method according to a nineteenth aspect extracts biometric features from biometric information, varies a positional relationship in a feature space between a representative feature that represents the biometric features of each individual, the representative feature being calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features, and calculates a loss function for training a model that extracts the biometric features based on a distance in the feature space between the representative feature and the biometric features of a first individual when the positional relationship is varied, and the distance between the biometric feature of the first individual and the representative feature of another individual.

[0086] (20) A storage medium according to a twentieth aspect is a storage medium storing a program for causing a computer to execute a process of extracting biometric features from biometric information, changing a positional relationship in a feature space between a representative feature representing the biometric features of each individual calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features, and updating parameters of a learning model used to extract the biometric features so that when the positional relationship is changed, the distance in the feature space between the biometric feature of a first individual and the representative feature of the first individual becomes smaller and the distance between the biometric feature of the first individual and the representative feature of another individual becomes larger.

[0087] (21) A storage medium according to a 21st aspect is a storage medium storing a program that causes a computer to execute a process of extracting biometric features from biometric information and generating a representative feature that represents a class of the biometric features for each individual, calculating a loss indicating the distance between the biometric feature extracted by the extraction means and the representative feature related to the biometric feature, and updating parameters of a learning model so as to reduce the loss, and in calculating the loss, setting a larger margin as the distance between the biometric feature of one individual and the representative feature of another individual is smaller based on the distance between the biometric feature of the one individual and the representative feature of the other individual.

[0088] (22) A storage medium according to a 22nd aspect is a storage medium storing a program for a computer to extract biometric features from biometric information, vary a positional relationship in a feature space between a representative feature that represents the biometric features of each individual calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features, and calculate a loss function for training a model that extracts the biometric features based on a distance in the feature space between the representative feature and the biometric features of a first individual when the positional relationship is varied, and the distance between the biometric feature of the first individual and the representative feature of another individual.

[0089] (23) A program according to a 23rd aspect is a program that causes a computer to execute a process of extracting biometric features from biometric information, changing a positional relationship in a feature space between a representative feature that represents the biometric features of each individual, the representative feature being calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features, and updating parameters of a learning model used to extract the biometric features so that, when the positional relationship is changed, a distance in the feature space between the biometric feature of a first individual and the representative feature of the first individual becomes smaller and a distance between the biometric feature of the first individual and the representative feature of another individual becomes larger.

[0090] (24) A program according to a 24th aspect causes a computer to execute a process of extracting biometric features from biometric information and generating a representative feature representing a class of the biometric features for each individual, calculating a loss indicating a distance between the biometric feature extracted by the extraction means and the representative feature related to the biometric feature, and updating parameters of a learning model so as to reduce the loss, and in calculating the loss, the program calculates the loss based on the distance between the representative feature and the biometric feature of a certain individual and the distance between the biometric feature of the certain individual and the representative feature of another individual, by setting a larger margin as the distance between the biometric feature of the certain individual and the representative feature of another individual is smaller.

[0091] (25) A program according to a 25th aspect is a recording medium storing a program for a computer to extract biometric features from biometric information, vary a positional relationship in a feature space between a representative feature that represents the biometric features of each individual calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features, and calculate a loss function for training a model that extracts the biometric features based on a distance in the feature space between the representative feature and the biometric features of a first individual when the positional relationship is varied, and a distance between the biometric feature of the first individual and the representative feature of another individual.

[0092] According to the above-described information processing system, information processing device, information processing method, and storage medium, it is possible to improve the discrimination performance of a learning model.

[0093] 100, 100e...Learning system, 11...Learning data acquisition means, 12...Iris feature extraction means, 13...Iris feature expansion means, 14...Loss function calculation means, 15...Gradient calculation means, 16...Parameter update means, 13a...Iris feature expansion means, 131a...Random number generation means, 132a...Iris feature conversion means, 13b...Iris feature expansion means, 131b...Representative feature storage means, 132b...Iris feature conversion means, 13c...Iris feature expansion means, 131c...Representative feature storage means, 132c...Iris feature conversion means, 13d...Representative feature expansion means, 131d...Representative feature storage means, 132d...Representative feature expansion means, 133d...Loss weight calculation means, 14d...Loss function calculation means, 14e...Loss function calculation means, 141e...Similarity calculation means, 142e...Margin calculation means, 143e...Loss calculation means

Claims

1. An information processing system comprising: an extraction means for extracting biometric features from biometric information; an extension means for changing a positional relationship in a feature space between a representative feature that represents the biometric features of each individual, the representative feature being calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information; and a parameter update means for updating parameters of a learning model that constitutes the extraction means so that when the positional relationship is changed, the distance in the feature space between the biometric feature of a first individual and the representative feature related to the first individual becomes smaller and the distance between the biometric feature of the first individual and the representative feature related to another individual becomes larger.

2. The information processing system of claim 1, further comprising: a loss function calculation means for calculating a loss indicating the difference in distance between the representative feature and the biometric feature of the first individual; and a gradient calculation means for calculating a gradient of the loss with respect to the parameters, wherein the parameter update means updates the parameters so as to minimize the difference in distance based on the gradient.

3. An information processing system as described in claim 1 or claim 2, comprising: a means for identifying the individual indicated by the biometric features based on the biometric features extracted by the extraction means and the representative feature for each individual; and a means for authenticating the person to be authenticated based on whether the identified individual matches the person to be authenticated.

4. The information processing system according to claim 1 or 2, wherein the extension means changes the biometric feature associated with a first individual to generate an extended biometric feature in which the distance from the representative feature associated with that individual is changed.

5. An information processing system according to claim 1 or claim 2, wherein the extension means generates extended biometric features in which the distance from the representative feature associated with the first individual is changed in various patterns by correcting the biometric feature associated with the first individual with a value generated from a random number.

6. The information processing system according to claim 1 or claim 2, wherein the extension means generates an extended biometric feature by rotating a vector indicated by the biometric feature relating to a first individual by a predetermined angle in a direction away from a vector indicated by the representative feature relating to said vector, thereby changing the extended biometric feature so that the distance from the representative feature increases.

7. The information processing system according to claim 1 or claim 2, wherein the extension means generates an extended biometric feature in which the distance from the representative feature of one of the individuals is increased by rotating the vector indicated by the biometric feature of a first of the individuals by a predetermined angle in a direction that brings the vector indicated by the representative feature of another of the individuals closer to the vector indicated by the representative feature of the other of the individuals.

8. An information processing system according to claim 1 or claim 2, wherein the extension means changes the representative feature associated with a first individual to generate an extended representative feature in which the distance from the biometric feature associated with that individual is changed.

9. An information processing system according to claim 1 or claim 2, wherein the extension means generates an extended representative feature in which the distance is changed by rotating a vector indicated by the representative feature related to a first said individual by a predetermined angle in a direction in which the vector is brought closer to a vector indicated by the representative feature related to another said individual.

10. The information processing system according to claim 9, further comprising: calculating the distance between the extended representative feature and the representative feature for all of the individuals; converting the calculated distance using a softmax function; setting the value relating to a first individual from among the calculated distances as a loss weight; and multiplying a predetermined loss function by the set loss weight to calculate a loss.

11. The information processing system of claim 2, wherein the loss function calculation means calculates the loss based on the distance between the representative feature and the biometric feature of a first individual and the distance between the biometric feature of the first individual and the representative feature of the other individual, by setting a larger margin value the smaller the distance between the biometric feature of the first individual and the representative feature of the other individual.

12. An information processing system comprising: an extraction means for extracting biometric features from biometric information and generating a representative feature representing a class of the biometric features for each individual; a loss function calculation means for calculating a loss indicating a distance between the biometric features extracted by the extraction means and the representative feature related to the biometric features; and a parameter update means for updating parameters of a learning model that constitutes the extraction means so as to reduce the loss, wherein the loss function calculation means calculates the loss based on the distance between the representative feature and the biometric feature of a first individual and the distance between the biometric feature of the first individual and the representative feature of another individual, by setting a larger value as a margin the smaller the distance between the biometric feature of the first individual and the representative feature of the other individual.

13. An information processing system comprising: an extraction means for extracting biometric features from biometric information; an extension means for changing a positional relationship in feature space between a representative feature that represents the biometric features of each individual, the representative feature being calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information; and a loss function calculation means for calculating a loss function for training a model that extracts the biometric features, based on the distance in the feature space between the representative feature and the biometric feature of a first individual when the positional relationship is changed, and the distance between the biometric feature of the first individual and the representative feature of another individual.

14. An information processing device comprising: an extraction means for extracting biometric features from biometric information; an extension means for changing a positional relationship in feature space between a representative feature that represents the biometric features of each individual, the representative feature being calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information; and a parameter update means for updating parameters of a learning model that constitutes the extraction means so that when the positional relationship is changed, the distance in the feature space between the biometric feature of a first individual and the representative feature relating to the first individual becomes smaller, and the distance between the biometric feature of the first individual and the representative feature relating to the other individuals becomes larger.

15. An information processing method comprising: extracting biometric features from biometric information; varying a positional relationship in a feature space between a representative feature that represents the biometric features of each individual, the representative feature being calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information; and updating parameters of a learning model used to extract the biometric features so that when the positional relationship is varied, the distance in the feature space between the biometric feature of a first individual and the representative feature relating to the first individual becomes smaller and the distance between the biometric feature of the first individual and the representative feature relating to the other individuals becomes larger.

16. A recording medium storing a program for causing a computer to execute the following processes: extracting biometric features from biometric information; changing the positional relationship in feature space between a representative feature that represents the biometric features of each individual, which is calculated based on correct answer information indicating which individual the biometric information belongs to and the biometric features extracted from the biometric information, and the biometric features; and updating parameters of a learning model used to extract the biometric features so that when the positional relationship is changed, the distance in feature space between the biometric feature of a first individual and the representative feature related to the first individual becomes smaller and the distance between the biometric feature of the first individual and the representative feature related to another individual becomes larger.

Images