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

Abstract

This information processing device comprises a first acquisition unit and a second acquisition unit. The first acquisition unit acquires a first simulated feature quantity based on a random number according to a precribed distribution. The second acquisition unit acquires simulated biological information based on the first simulated feature quantity using a biological information generation model for generating, from the first simulated feature quantity, simulated biological information modeled after a type of biological information determined in advance.

Description

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

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

[0002] For example, Patent Document 1 discloses a search system that generates new item candidates using a person's face image (a two-dimensional still image) as an item.

[0003] The search system described in Patent Literature 1 outputs to a user terminal a first item having a first component value and multiple item candidates having component values ​​different from the first component value for multiple principal components that make up an item. The search system identifies a second item selected on the user terminal from the item candidates and calculates the positional relationship between the first component value and the second component value of the second item for each principal component. The search system calculates a component value distribution for each principal component according to the positional relationship, and newly generates multiple item candidates for the second item based on the component value distribution and outputs them to the user terminal.

[0004] The initial image of the first item (reference image) is an average item (average face) based on the average value of each principal component calculated by principal component analysis of the feature quantities of the face image serving as the teacher image. Here, a known machine learning image generation technique is used.

[0005] Japanese Patent Application Laid-Open No. 2023-002325

[0006] The present disclosure aims to improve upon the techniques described in the prior art documents mentioned above.

[0007] The information processing device of the present disclosure includes a first acquisition means for acquiring a first simulated feature based on a random number according to a predetermined distribution, and a second acquisition means for acquiring the simulated biometric information based on the first simulated feature using a biometric information generation model that generates simulated biometric information that imitates a predetermined type of biometric information from the first simulated feature.

[0008] The information processing system of the present disclosure includes a first acquisition means for acquiring a first simulated feature based on a random number according to a predetermined distribution, and a second acquisition means for acquiring the simulated biometric information based on the first simulated feature using a biometric information generation model that generates simulated biometric information that imitates a predetermined type of biometric information from the first simulated feature.

[0009] The information processing method disclosed herein includes one or more computers acquiring first simulated features based on random numbers following a predetermined distribution, and acquiring the simulated biometric information based on the first simulated features using a biometric information generation model that generates simulated biometric information that imitates a predetermined type of biometric information from the first simulated features.

[0010] A recording medium according to the present disclosure is a recording medium having recorded thereon a program for causing one or more computers to acquire first simulated features based on random numbers according to a predetermined distribution, and to acquire simulated biometric information based on the first simulated features using a biometric information generation model that generates simulated biometric information that imitates a predetermined type of biometric information from the first simulated features.

[0011] 1 is a block diagram showing an example configuration of a first information processing device according to the present disclosure. FIG. 2 is a flowchart showing processing operations of the first information processing device according to the present disclosure. FIG. 3 is a diagram showing an example of a correspondence relationship between a feature space representing a feature vector and an iris image space representing a circular iris image. FIG. 4 is a diagram showing an example of a correspondence relationship between a feature space representing a feature map and an iris image space representing a rectangular iris image. FIG. 5 is a block diagram showing a physical configuration of a first information processing device according to the present disclosure. FIG. 6 is a block diagram showing an example configuration of a second information processing device according to the present disclosure. FIG. 7 is a flowchart showing processing operations of the second information processing device according to the present disclosure. FIG. 8 is a block diagram showing an example configuration of a first update unit according to the present disclosure. FIG. 9 is a flowchart showing processing operations of the first update unit according to the present disclosure. FIG. 10 is a block diagram showing an example configuration of a third information processing device according to the present disclosure. FIG. 11 is a flowchart showing processing operations of the third information processing device according to the present disclosure. FIG. 12 is a block diagram showing an example configuration of a first likelihood model update unit according to the present disclosure. FIG. 13 is a flowchart showing processing operations of the first likelihood model update unit according to the present disclosure. FIG. 14 is a block diagram showing an example configuration of a fourth information processing device according to the present disclosure. FIG. 15 is a flowchart showing processing operations of the fourth information processing device according to the present disclosure. FIG. 16 is a block diagram showing an example configuration of a first local evaluation unit according to the present disclosure. FIG. 17 is a flowchart showing processing operations of the first local evaluation unit according to the present disclosure. FIG. 18 is a block diagram showing an example configuration of a second update unit according to the present disclosure. 10 is a flowchart showing the processing operation of a second update unit according to the present disclosure. FIG. 11 is a block diagram showing a configuration example of a fifth information processing device according to the present disclosure. FIG. 12 is a flowchart showing the processing operation of the fifth information processing device according to the present disclosure. FIG. 13 is a block diagram showing a configuration example of a first determination unit according to the present disclosure. FIG. 14 is a flowchart showing the processing operation of the first determination unit according to the present disclosure. FIG. 15 is a block diagram showing a configuration example of a sixth information processing device according to the present disclosure. FIG. 16 is a flowchart showing the processing operation of the sixth information processing device according to the present disclosure. FIG. 17 is a block diagram showing a configuration example of a third update unit according to the present disclosure. FIG. 18 is a flowchart showing the processing operation of the third update unit according to the present disclosure. FIG. 19 is a block diagram showing a configuration example of a seventh information processing device according to the present disclosure. FIG. 19 is a flowchart showing the processing operation of the seventh information processing device according to the present disclosure.Fig. 10 is a block diagram illustrating a configuration example of an eighth information processing apparatus according to the present disclosure. Fig. 11 is a flowchart illustrating a processing operation of the eighth information processing apparatus according to the present disclosure. Fig. 12 is a block diagram illustrating a configuration example of a first information processing system according to the present disclosure.

[0012] Hereinafter, in this disclosure, the drawings relate to one or more embodiments. In addition, in all drawings, similar components are given similar reference numerals and descriptions thereof will be omitted as appropriate.

[0013] [First Embodiment] (Overview) In biometric authentication, features that can identify an individual are extracted from biometric information including irises, faces, fingerprints, veins, etc., and authentication is performed using these features. When a machine learning model is used to extract features from biometric information, biometric information from many individuals is used as training data to improve the accuracy of biometric authentication. However, it can be difficult to obtain biometric information from many real individuals.

[0014] As described above, Patent Document 1 discloses a technique for generating a plurality of new item candidates based on the component value distribution for a second item selected by a user. Patent Document 1 also discloses the use of a person's face image as an item.

[0015] The multiple face images newly generated using the technology described in Patent Literature 1 can be referred to as virtual individual face images, i.e., simulated face images that imitate face images. However, the simulated face images generated using the technology described in Patent Literature 1 are likely to have biased characteristics depending on the user's selection, training images, etc. Biometric authentication typically targets many individuals with a wide variety of facial expressions, etc. Therefore, if the simulated face images generated using the technology described in Patent Literature 1 are used as training data, the accuracy of biometric authentication may not be sufficiently improved.

[0016] One of the objectives of the present disclosure is to generate simulated biometric information with various variations belonging to many individuals.

[0017] As shown in FIG. 1 , the information processing device 100 according to this embodiment includes a first acquisition unit 110 and a second acquisition unit 120 .

[0018] The first acquiring unit 110 acquires a first simulated feature based on random numbers according to a predetermined distribution.

[0019] The second acquisition unit 120 acquires simulated biometric information based on the first simulated feature using a biometric information generation model that generates simulated biometric information that imitates a predetermined type of biometric information from the first simulated feature.

[0020] According to the information processing device 100, simulated biometric information can be generated using random numbers that follow a predetermined distribution, making it possible to generate simulated biometric information with various variations that belong to many individuals.

[0021] The information processing device 100 executes information processing as shown in FIG.

[0022] The first acquisition unit 110 acquires a first simulated feature based on random numbers according to a predetermined distribution (step S110).

[0023] The second acquisition unit 120 acquires simulated biometric information based on the first simulated feature using a biometric information generation model that generates simulated biometric information that imitates a predetermined type of biometric information from the first simulated feature (step S120).

[0024] This information processing allows the generation of simulated biometric information using random numbers that follow a predetermined distribution, making it possible to generate simulated biometric information with a wide variety of variations that belong to many individuals.

[0025] (Detailed Example) "Biometric information" is information obtained from a living body (biometric information derived from a living body) or information that imitates this (simulated biometric information). The type of biometric information may be determined in advance.

[0026] When biometric information is used for authentication, the biometric information may be, for example, an image including a specific part of a living body (biometric image). The specific part may be, for example, a part that can identify an individual and is used for authentication.

[0027] The type of biometric information may be determined according to a specific body part, such as an iris image used for iris authentication, a face image used for face authentication, a fingerprint image used for fingerprint authentication, a vein image used for vein authentication, or a palm image used for palm authentication.

[0028] The iris image may include the iris and may further include all or a part of the body, such as the eye or face. The face image may include the face and may further include all or a part of the body including the face. The fingerprint image may include a fingerprint pattern and may further include all or a part of the body including the fingerprint. The vein image may include a vein pattern and may further include all or a part of the body including the veins. The palm image may include a palm print or the like and may further include all or a part of the body other than the palm.

[0029] The types of biometric information are not limited to those exemplified here.

[0030] The biological information derived from the living body is obtained, for example, using a sensor device (not shown). The sensor device may be, for example, a camera that uses light in an appropriate wavelength band, such as the visible light region or the near-infrared light region. The sensor device may read a fingerprint image including a fingerprint pattern using a capacitive electrostatic method or an ultrasonic method.

[0031] The living body may be a human, or may be an animal such as a dog, cat, cow, pig, bird, snake, etc.

[0032] In the following, an example will be described in which the living body is a human, and an example will be described in which the biometric information and simulated biometric information are biometric images including a specific body part.

[0033] When the information processing device 100 uses biometric information derived from a living organism, the information processing device 100 may be equipped with a sensor device or the like, or the information may be acquired from the sensor device or the like via a communication network or the like configured using a wired, wireless, or combination thereof.

[0034] The method for obtaining biological information derived from a living body is not limited to the example given here.

[0035] The simulated biometric information does not have to be biometric information obtained from an actual living body, and can also be called biometric information of a virtual individual.

[0036] The "feature amount" is information extracted from biometric information. The feature amount has the ability to identify an individual.

[0037] For example, in biometric authentication, features are extracted from biometric information using an extraction model to identify an individual. The extraction model is, for example, a machine learning model that extracts features from biometric information. The extraction model outputs features when biometric information is input. The extraction model may be, for example, a deep learning model configured with a neural network.

[0038] In a learning-based extraction model, for example, training data including training biometric information and a correct label indicating a class to which the training biometric information belongs is used. The training biometric information is biometric information prepared in advance for learning. The extraction model during training outputs training features, which are features based on the training biometric information, in response to input of the training biometric information. One or more extraction model parameters included in the extraction model are updated so that the training features are classified into a correct label indicating a class to which they belong.

[0039] The method for extracting features, configuring an extraction model, and building an extraction model are not limited to the methods exemplified here, and general methods may be used, for example.

[0040] The feature amount may be represented, for example, by an N-dimensional tensor, where N is an integer equal to or greater than 1. A feature amount represented by a one-dimensional tensor (vector) is generally also referred to as a feature vector. A feature amount represented by a two-dimensional tensor is also referred to as a "feature map."

[0041] The feature amount may be represented by a tensor, for example, where at least one of the dimension and the norm is 1. This makes it possible to reduce the processing load when using the feature amount.

[0042] Here, for example, when the biometric information is an iris image, a feature map and a rectangular iris image may be used for the feature amount and the iris image, respectively.

[0043] A "circular iris image" is an image showing an iris in the shape of a circular band, such as an iris shown in an iris image obtained by general photography, etc. A "rectangular iris image" is an image obtained by converting a circular iris image into a rectangle. A rectangular iris image has sides corresponding to the radial and circumferential directions of the iris.

[0044] That is, the predetermined type of biometric information may be an image showing an iris, and may include an iris image showing the iris in a circular band shape, or may be a rectangular iris image having sides corresponding to the radial and circumferential directions of the iris.

[0045] When converting a circular iris region into a rectangular iris image, the conversion process may include, for example, conversion from a polar coordinate system to a Cartesian coordinate system, and if there are pixels that do not correspond before and after the conversion, the pixel values ​​may be found using linear interpolation, nearest neighbor interpolation, bilinear interpolation, bicubic interpolation, area interpolation, etc. Note that the process of converting into a rectangular iris image or the like is not limited to the examples given here, and general processing techniques may also be used.

[0046] 3 is a diagram showing an example of a correspondence relationship between a feature space representing a feature vector and an iris image space representing a circular iris image, and FIG. 4 is a diagram showing an example of a correspondence relationship between a feature space representing a feature map and an iris image space representing a rectangular iris image.

[0047] In a pair of a feature vector and a circular iris image, it is unknown which partial feature in the feature vector represents which feature of the pattern in the iris image. Therefore, the feature vector must simultaneously represent the iris pattern and its position, and this relationship is generally complex.

[0048] On the other hand, a pair of a feature map and a rectangular iris image is geometrically constrained by having two sides corresponding to the radial and circumferential directions of the iris, and the positions of partial features in the feature map generally coincide with the positions of pattern features in the rectangular iris image used to extract the feature map. Therefore, a partial feature at a certain position in the feature map only needs to represent the local feature at the corresponding position in the rectangular iris image, and does not need to represent the position information of the feature.

[0049] The positional correspondence between partial features in a feature map and partial images in a rectangular iris image is simpler than the positional correspondence between partial features in a feature vector and partial images in a circular iris image. Therefore, using a feature map and a rectangular iris image for each feature and iris image allows for easier mapping than using a feature vector and a circular iris image for each feature and iris image. For example, a simulated rectangular iris image, which is simulated biometric information simulating a rectangular iris image, can be easily generated from a first simulated feature in feature map format (two-dimensional map format). For example, a feature map can be easily extracted from a rectangular iris image. For example, a second simulated feature map, which is a two-dimensional map format feature, can be easily extracted from the simulated rectangular iris image.

[0050] (Regarding the first acquisition unit 110) The first acquisition unit 110 acquires random numbers that follow a predetermined distribution, for example, in response to a user instruction or the like. The random numbers are generated, for example, by a random number generator. The random number generator may be included in the first acquisition unit 110, or may be provided in an external device (not shown). When the external device has a random number generator, the first acquisition unit 110 may acquire the random numbers generated by the external device via a communication network or the like, for example.

[0051] Examples of the predetermined distribution include uniform distribution, normal distribution, Poisson distribution, geometric distribution, exponential distribution, gamma distribution, t distribution, chi-squared distribution, and F distribution. However, the predetermined distribution is not limited to these. The random numbers generated by the random number generator may be pseudo-random numbers or true random numbers.

[0052] The first acquiring unit 110 generates a first simulated feature using, for example, the acquired random number.

[0053] The first simulated feature is generated using random numbers and is expressed in a predetermined format, such as an N-dimensional tensor, where N is a predetermined integer greater than or equal to 1.

[0054] In more detail, for example, the first simulated feature may be a first simulated feature vector represented by a one-dimensional tensor (vector). Alternatively, for example, the first simulated feature may be a first simulated feature map represented by a two-dimensional tensor. Such a first simulated feature may be, for example, a first simulated feature vector or a first simulated feature map in which generated random numbers are assigned to each component.

[0055] More specifically, for example, the first simulated feature may be a first simulated feature vector or a first simulated feature map expressed as a vector or a tensor composed of N×M components. M is a predetermined integer equal to or greater than 1. When N is 1, the first simulated feature is a 1×M first simulated feature vector. When N is 2 or greater, the first simulated feature is an N×M first simulated feature map. The first acquisition unit 110 may acquire, for example, a random number sequence composed of N×M random numbers. The first acquisition unit 110 may sequentially arrange the components of the random number sequence into each component of a vector or tensor according to a predetermined arrangement order. As a result, the first acquisition unit 110 generates a first simulated feature vector or a first simulated feature map. That is, the first simulated feature may include a first simulated feature vector in the form of a vector (one-dimensional tensor) in which random numbers are arranged as components of a vector. The first simulated feature may include a feature (first simulated feature map) in a two-dimensional map format in which random numbers are arranged as components of a two-dimensional tensor.

[0056] Assume that the first simulated feature is represented by a first simulated feature map. Furthermore, assume that the simulated biometric information acquired based on the first simulated feature map includes an image including a biometric image, i.e., a simulated biometric image simulating a biometric image. Specifically, for example, assume that the simulated biometric information includes a simulated rectangular iris image simulating a rectangular iris image. In this case, the aspect ratios of the first simulated feature map and the simulated rectangular iris image may be the same. This facilitates associating the positional relationship between the first simulated feature map and the simulated rectangular iris image, thereby facilitating mapping between the first simulated feature map and the simulated rectangular iris image. As a result, a higher-quality simulated biometric image can be acquired based on the first simulated feature map.

[0057] Note that the method for generating the first simulated feature using a random number is not limited to the method exemplified here. The first simulated feature may also be generated by an external device (not shown). In this case, the first acquisition unit 110 may acquire the first simulated feature generated by the external device via a communication network or the like. This external device may be the same as or different from the device including the random number generator described above.

[0058] (Regarding the second acquisition unit 120) The second acquisition unit 120 acquires simulated biometric information based on the first simulated feature by, for example, acquiring the first simulated feature and inputting it into a biometric information generation model. The second acquisition unit 120 may include, for example, the biometric information generation model. The biometric information generation model is a model that outputs simulated biometric information when the first simulated feature is input.

[0059] The biometric information generation model may be, for example, a deep learning model using a convolutional neural network, an attention mechanism, or the like. The biometric information generation model may be, for example, a generative model trained using adversarial learning such as GAN (Generative Adversarial Networks). The biometric information generation model may be, for example, a stochastic diffusion model. When using a stochastic diffusion model, the biometric information generation model learns the distribution of training biometric information and sequentially maps a first simulated feature map to the learned distribution, thereby sequentially generating simulated biometric images from the first simulated feature map.

[0060] The learning biometric information is biometric information prepared for learning.

[0061] The method for constructing a biometric information generation model is not limited to the method described here or in other embodiments.

[0062] The predetermined type of biometric information may be, for example, an iris image, a face image, a fingerprint image, a vein image, etc. If the biometric information is an iris image, the predetermined type of biometric information may be, for example, a circular iris image that imitates a circular iris image, or a rectangular iris image that imitates a rectangular iris image. Note that the predetermined type of biometric information is not limited to those exemplified here.

[0063] (Example of physical configuration of information processing device 100) The information processing device 100 physically includes a bus 1010, a processor 1020, a memory 1030, a storage device 1040, a network interface 1050, an input interface 1060, and an output interface 1070, as shown in FIG.

[0064] The bus 1010 is a data transmission path for transmitting and receiving data among the processor 1020, memory 1030, storage device 1040, network interface 1050, input interface 1060, and output interface 1070. However, the method of connecting the processor 1020 and the like to each other is not limited to bus connection.

[0065] The processor 1020 is implemented by a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or the like.

[0066] The memory 1030 is a main storage device realized by a RAM (Random Access Memory) or the like.

[0067] The storage device 1040 is an auxiliary storage device realized by a hard disk drive (HDD), a solid state drive (SSD), a memory card, a read only memory (ROM), or the like. The storage device 1040 stores program modules for realizing the functions of the information processing device 100. The processor 1020 reads each of these program modules into the memory 1030 and executes them to realize the function corresponding to the program module.

[0068] The network interface 1050 is an interface for connecting the information processing device 100 to a network. The network is a communication network for transmitting and receiving information to and from other devices (not shown), and may be wired, wireless, or a combination of these.

[0069] The input interface 1060 is an interface for the user to input information, and is composed of, for example, a touch panel, a keyboard, a mouse, and the like.

[0070] The output interface 1070 is an interface for presenting information to the user, and is configured, for example, by a liquid crystal panel, an organic EL (Electro-Luminescence) panel, or the like.

[0071] The information processing device 100 may be physically configured as a plurality of devices communicably connected to each other via a wired, wireless, or combination thereof communication network, etc. In this case, each of the plurality of devices may be physically configured in the same manner as the information processing device 100, for example.

[0072] (Actions and Effects) As described above, according to this embodiment, the information processing device 100 includes the first acquisition unit 110 and the second acquisition unit 120. The first acquisition unit 110 acquires first simulated features using random numbers according to a predetermined distribution. The second acquisition unit 120 acquires simulated biometric information based on the first simulated features using a biometric information generation model that generates simulated biometric information that simulates predetermined types of biometric information from the first simulated features.

[0073] This allows simulated biometric information to be generated using random numbers that follow a predetermined distribution, making it possible to generate simulated biometric information with a wide variety of variations that belong to many individuals.

[0074] For example, by using simulated biometric information to train a machine learning model used in biometric authentication, it becomes possible to improve the accuracy of biometric authentication. Note that this effect is not intended to limit the use of simulated biometric information to biometric authentication.

[0075] According to this embodiment, the predetermined type of biometric information includes an image of an iris, the image being a rectangular iris image having sides corresponding to the radial and circumferential directions of the iris, and the first simulated feature includes a feature in a two-dimensional map format in which random numbers are arranged as components of a two-dimensional tensor.

[0076] This makes it possible to easily generate a simulated rectangular iris image from the first simulated feature amount in the form of a two-dimensional map, thereby making it possible to easily generate simulated biometric information.

[0077] According to this embodiment, the first simulated feature is expressed as a tensor having at least one of a dimension and a norm of one.

[0078] This reduces the processing load using feature amounts, and therefore makes it possible to easily generate simulated biometric information with a wide variety of variations belonging to many individuals.

[0079] Second Embodiment In this embodiment, an example of a method for constructing a biometric information generation model will be described. Note that descriptions that overlap with other embodiments will be omitted as appropriate for the sake of brevity.

[0080] As shown in FIG. 6, the information processing apparatus 200 according to this embodiment includes a first acquisition unit 110 and a second acquisition unit 120 similar to those in the first embodiment, and a learning unit 230.

[0081] The learning unit 230 learns the biometric information generation model.

[0082] In detail, for example, the learning unit 230 includes a biological likelihood estimating unit 231 and an updating unit 232 .

[0083] The biometric likelihood estimation unit 231 estimates a simulated biometric likelihood, which is a biometric likelihood based on simulated biometric information, using a biometric likelihood estimation model that estimates a biometric likelihood indicating the likelihood that the biometric information is of a predetermined type.

[0084] The update unit 232 updates one or more model parameters included in the biometric information generation model using the simulated biometric likelihood.

[0085] The information processing device 200 according to the present disclosure executes information processing as shown in FIG.

[0086] Steps S110 and S120 are executed in the same manner as in the first embodiment.

[0087] The learning unit 230 learns the biometric information generation model (step S230).

[0088] In detail, for example, the learning unit 230 executes the following learning process (step S230).

[0089] The biometric likelihood estimation unit 231 estimates a simulated biometric likelihood, which is a biometric likelihood based on simulated biometric information, using a biometric likelihood estimation model that estimates a biometric likelihood indicating the likelihood that the biometric information is a predetermined type of biometric information (step S231).

[0090] The update unit 232 updates one or more model parameters included in the biometric information generation model using the simulated biometric likelihood (step S232).

[0091] (Regarding the biological likelihood estimation unit 231) The biological likelihood estimation unit 231 acquires, for example, simulated biological information based on the first simulated feature acquired by the second acquisition unit 120 using a biological information generation model currently being trained. The biological likelihood estimation unit 231 estimates a simulated biological likelihood by, for example, inputting the simulated biological information into a biological likelihood estimation model.

[0092] The simulated biometric likelihood is a biometric likelihood for the simulated biometric information. The biometric likelihood is a value indicating the likelihood that the biometric information is a predetermined type of biometric information. In other words, when the biometric information input into the biometric likelihood estimation model is simulated biometric information, the simulated biometric likelihood is a value indicating the likelihood that the simulated biometric information is a predetermined type of biometric information.

[0093] The biological likelihood estimation unit 231 may include, for example, a trained biological likelihood estimation model that outputs a biological likelihood when biological information is input.

[0094] The biological likelihood estimation model may be, for example, a deep learning model configured with a convolutional neural network, etc. The biological likelihood estimation model may be, for example, a model trained using adversarial learning such as GAN to distinguish whether an input image is biological information derived from a living organism or simulated biological information.

[0095] The method for constructing a biological likelihood estimation model is not limited to this, and an example thereof will be described in another embodiment.

[0096] (Regarding the Updater 232) As shown in FIG. 8, the updater 232 includes a first loss calculator 232a and a parameter updater 232b.

[0097] The first loss calculation unit 232a calculates the first loss using the simulated living body likelihood.

[0098] The parameter update unit 232b uses the first loss to update one or more model parameters included in the biological information generation model.

[0099] The update unit 232 executes an update process (step S232) as shown in FIG. 9, for example.

[0100] The first loss calculation unit 232a calculates the first loss using the simulated living body likelihood (step S232a).

[0101] The parameter update unit 232b uses the first loss to update one or more model parameters included in the biometric information generation model (step S232b).

[0102] (First Loss Calculation Unit 232a) The first loss calculation unit 232a calculates a mean square error as the first loss, for example, by using a predetermined value for a predetermined type of biometric information and a simulated biometric likelihood.

[0103] In detail, for example, it is assumed that the predetermined type of biometric information is an iris image, and the simulated biometric information is a simulated iris image that imitates the iris image. In this case, the simulated biometric likelihood z ｉ is the simulated iris image x input to the biometric likelihood estimation model. ｉ is a value that represents the degree to which the image resembles a real iris image. The simulated biological likelihood is z ｉ For example, it is a value between 0 and a specified value, and the specified value is assumed to be 1. Here, i is an integer between 1 and B, inclusive. B represents, for example, a batch size in mini-batch learning.

[0104] In this case, the first loss may be a loss expressed by the following equation (1) using, for example, a mean square error.

[0105] The first loss is not limited to the above example, and may be, for example, a mean absolute error or the like.

[0106] (Regarding the parameter update unit 232b) The parameter update unit 232b updates one or more model parameters included in the bioinformation generation model using a general optimization method such as a search method such as a gradient method or a grid search. Note that the method for updating the model parameters is not limited to these examples.

[0107] In more detail, for example, the gradient method is a method in which the gradient of a parameter to be updated is calculated from the loss and the parameter is updated using the calculated gradient. When mini-batch learning is performed, the parameter update unit 232b may calculate the gradient of a parameter from the loss for each mini-batch and update the parameter using the calculated gradient.

[0108] (Actions and Effects) As described above, according to this embodiment, the information processing device 200 includes a learning unit 230 that learns a biometric information generation model. The learning unit 230 includes a biometric likelihood estimation unit 231 and an update unit 232. The biometric likelihood estimation unit 231 estimates a simulated biometric likelihood, which is a biometric likelihood based on the simulated biometric information, using a biometric likelihood estimation model that estimates a biometric likelihood indicating the likelihood that the biometric information is a predetermined type of biometric information. The update unit 232 updates one or more model parameters included in the biometric information generation model using the simulated biometric likelihood.

[0109] This makes it possible to generate simulated biometric information that resembles a predetermined type of biometric information. Therefore, it becomes possible to generate simulated biometric information that resembles a predetermined type of biometric information.

[0110] Third Embodiment In this embodiment, another example of a method for constructing a biological likelihood estimation model will be described. Note that descriptions that overlap with other embodiments will be omitted as appropriate for the sake of brevity.

[0111] As shown in FIG. 10, the information processing device 300 according to this embodiment includes a first acquisition unit 110 and a second acquisition unit 120 similar to those in the first embodiment, a learning unit 330, and a likelihood model learning unit 340.

[0112] The learning unit 330 learns the biometric information generation model.

[0113] The learning unit 330 includes a biological likelihood estimation unit 331 and an update unit 232 .

[0114] The biometric likelihood estimation unit 331 estimates a simulated biometric likelihood, which is a biometric likelihood based on simulated biometric information, using a biometric likelihood estimation model that estimates a biometric likelihood indicating the likelihood that the biometric information is of a predetermined type.

[0115] In detail, for example, the biometric likelihood estimation unit 331 estimates a simulated biometric likelihood, which is a biometric likelihood based on the simulated biometric information, using a trained biometric likelihood estimation model. The biometric likelihood estimated by the trained biometric likelihood estimation model indicates the likelihood that the biometric information belongs to a biometric information class among multiple classes including a biometric information class corresponding to a predetermined type of biometric information.

[0116] The likelihood model learning unit 340 learns the biological likelihood estimation model.

[0117] In detail, for example, the likelihood model learning unit 340 includes a likelihood model updating unit 341 and a training data storage unit 342 .

[0118] The likelihood model update unit 341 updates one or more model parameters included in the biological likelihood estimation model using the biological likelihood estimated by the biological likelihood estimation model during training.

[0119] The training data storage unit 342 stores training data, which is data prepared in advance for training the biological likelihood estimation model.

[0120] The information processing device 300 executes information processing as shown in FIG.

[0121] Steps S110 and S120 are executed in the same manner as in the first embodiment.

[0122] The learning unit 330 learns the biometric information generation model (step S330).

[0123] In detail, for example, the learning unit 330 executes the following learning process (step S330).

[0124] The biometric likelihood estimation unit 331 estimates a simulated biometric likelihood, which is a biometric likelihood based on simulated biometric information, using a biometric likelihood estimation model that estimates a biometric likelihood indicating the likelihood that the biometric information is a predetermined type of biometric information (step S331a).

[0125] In detail, for example, in step S331a, the biological likelihood estimation unit 331 uses the trained biological likelihood estimation model to estimate a simulated biological likelihood, which is a biological likelihood based on the simulated biological information.

[0126] The same update process (step S232) as in the second embodiment is executed.

[0127] Furthermore, the information processing device 300 performs learning of a biological likelihood estimation model (likelihood model learning process). By performing the likelihood model learning process, a learned biological likelihood estimation model is constructed. The above-described information processing (see FIG. 11 ) may be executed, for example, using the learned biological likelihood estimation model. As a result, in step S331a, a biological likelihood indicating the likelihood that the biological information belongs to a biological information class among a plurality of classes including a biological information class corresponding to a predetermined type of biological information is estimated.

[0128] The likelihood model learning process will be described in detail later.

[0129] (Regarding the Likelihood Model Updater 341) As shown in FIG. 12, the likelihood model updater 341 includes a likelihood loss calculator 341a and a likelihood parameter updater 341b, for example.

[0130] The likelihood loss calculation unit 341a calculates the likelihood loss using a training organism likelihood (described in detail later).

[0131] The likelihood parameter update unit 341b updates one or more model parameters included in the biological likelihood estimation model using the likelihood loss.

[0132] The information processing device 300 executes a likelihood model learning process as shown in FIG. 13, for example.

[0133] The biometric likelihood estimation unit 331 estimates a training biometric likelihood, which is a biometric likelihood based on likelihood learning information (described in detail below), using a biometric likelihood estimation model currently being trained to estimate a biometric likelihood indicating the likelihood that the biometric information is a predetermined type of biometric information (step S331b).

[0134] The likelihood loss calculation unit 341a calculates the likelihood loss using the training organism likelihood (step S341a).

[0135] The likelihood parameter update unit 341b updates one or more model parameters included in the biological likelihood estimation model using the likelihood loss (step S341b).

[0136] (Regarding Training Data) The training data includes, for example, likelihood learning information and a correct answer of the biological likelihood related to the likelihood learning information.

[0137] The likelihood learning information includes learning biometric information and learning information.

[0138] The training biometric information is a predetermined type of biometric information, and is prepared in advance for training a biometric likelihood estimation model.

[0139] The learning information is information that is different from the predetermined types of biometric information.

[0140] For example, if the predetermined type of biometric information is a human iris image, the training biometric information is a human iris image prepared in advance. In this case, the training information is, for example, a general image other than an iris. In detail, for example, the training information may include one or more of an image including a part or all of an animal, an image including a part or all of an object other than an animal, etc. Examples of animals include, but are not limited to, dogs, cats, cows, pigs, birds, and snakes. Examples of objects other than animals may include, for example, artificial objects, plants, mushrooms, food, etc. Examples of artificial objects may include, for example, buildings such as houses, industrial products such as automobiles, etc. Examples of plants may include, for example, flowers such as roses and tulips, and trees such as cedars and cypress, etc. Examples of food may include beef jerky, raw meat, etc. Note that the objects other than animals are not limited to those exemplified here.

[0141] The images of animals and non-animal objects may include, for example, cross-sections of raw meat, beef jerky, mushrooms, dog irises, etc., which have textures similar to human iris images.

[0142] The correct answer to the biological likelihood may be, for example, a correct label indicating the class to which the training information belongs.

[0143] (Regarding the biological likelihood estimation unit 331) The biological likelihood estimation unit 331 acquires, for example, training data from the training data storage unit 342. The biological likelihood estimation unit 331 inputs, for example, likelihood learning information included in the training data to a biological likelihood estimation model being trained, and estimates a training biological likelihood that is a biological likelihood based on the likelihood learning information.

[0144] The training data storage unit 342 may be provided in an external device (not shown). In this case, for example, the biological likelihood estimation unit 331 may acquire data from the external device via a wired, wireless, or a combination thereof communication network, instead of the training data storage unit 342.

[0145] (Regarding the Likelihood Loss Calculation Unit 341a) The likelihood loss calculation unit 341a calculates the likelihood loss using, for example, the training organism likelihood and the correct answer of the organism likelihood.

[0146] In more detail, for example, it is assumed that the predetermined type of biometric information is an iris image, and the likelihood training information is a likelihood training image. The likelihood training image here is an iris image prepared in advance for training the biometric likelihood estimation model. Examples of likelihood loss in this case include the following examples 1 and 2. Likelihood loss examples 1 and 2 are examples of likelihood loss that differ depending on the class into which the biometric likelihood estimation model classifies input biometric information (e.g., simulated biometric information).

[0147] (Example 1 of likelihood loss) The correct answer for biometric likelihood may be, for example, a correct label indicating a first class to which an iris image (i.e., biometric information for training) belongs and a second class to which a non-iris image, which is an image other than an iris image (i.e., biometric information for training), belongs.

[0148] For example, the likelihood learning image x input to the biological likelihood estimation model ｉ The correct label y ｉ is set to "1" or "0". Here, i is an integer between 1 and B, inclusive. B represents, for example, the batch size in mini-batch learning. Likelihood learning image x ｉ For example, the probability of belonging to the first class (training likelihood) is z ｉ and the probability of belonging to the second class is 1-z ｉ Let us assume that:

[0149] In this case, the likelihood loss may be expressed by the following equation (2) using, for example, binary cross-entropy loss.

[0150] (Example 2 of Correct Answer for Biometric Likelihood) The correct answer for biometric likelihood may be, for example, a correct label indicating each of the first to Nth classes. K is an integer equal to or greater than 2. For example, one of the first to Nth classes (for example, the first class) is a class to which an iris image (i.e., biometric information for training) belongs, and the other classes are classes to which likelihood training images (i.e., training information) other than iris images belong.

[0151] Likelihood learning image x input to the biological likelihood estimation model ｉ The correct label y ｊ ｉ is the likelihood learning image x ｉ The value of the class to which y belongs is 1, and other values ​​are 0. Here, i is an integer between 1 and B, inclusive. B represents, for example, the batch size in mini-batch learning. j is an integer between 1 and N, inclusive. N here is the number of classes. Correct label y ｊ ｉ For example, the correct answer label y ｊ ｉ is expressed as a vector containing the correct answer of the jth class in the jth component, and if the number of classes N is 3, then the likelihood learning image x ｉ The correct label y ｉ is represented as (0,0,1).

[0152] Also, the likelihood learning image x ｉ For example, the probability of belonging to the jth class (training likelihood) is z ｊ ｉ Let us assume that:

[0153] In this case, the likelihood loss may be expressed by the following equation (3) using, for example, binary cross-entropy loss.

[0154]

[0155] Note that the likelihood loss is not limited to the above example.

[0156] (Regarding the likelihood parameter update unit 341b) The likelihood parameter update unit 341b updates one or more model parameters included in the biometric information generation model so as to reduce likelihood loss. In detail, for example, the likelihood parameter update unit 341b updates one or more model parameters included in the biometric likelihood estimation model using a general optimization method such as a search method such as a gradient method or a grid search. Note that the method of updating the model parameters is not limited to the example given here.

[0157] In more detail, for example, the gradient method is a method in which the gradient of a parameter to be updated is calculated from the likelihood loss and the parameter is updated using the calculated gradient. When mini-batch learning is performed, the likelihood parameter update unit 341b may calculate the gradient of a parameter from the likelihood loss for each mini-batch and update the parameter using the calculated gradient.

[0158] According to this embodiment, a biometric likelihood estimation model is trained using training biometric information, which is a predetermined type of biometric information, and training information different from the predetermined type of biometric information. Therefore, when a class to which the training biometric information belongs is defined as a biometric information class, it is possible to estimate, as a biometric likelihood, the likelihood that a biometric image input to the biometric likelihood estimation model belongs to the biometric information class among multiple classes including the biometric information class.

[0159] For example, the predetermined type of biometric information is an iris image, and the biometric information class to which the iris image belongs is the iris image class. Furthermore, assume that the learning information includes an image of a dog, an image of a cat, and an image of a cross-section of meat. The classes to which the image of the dog, the image of the cat, and the image of the cross-section of meat belong are the dog class, the cat class, and the meat cross-section class, respectively. In this case, the biometric likelihood estimation model can estimate, as the biometric likelihood, the likelihood that the biometric image input to the biometric likelihood estimation model belongs to the iris image class among the iris image class, the dog class, the cat class, and the meat cross-section class.

[0160] Generally, when using adversarial learning such as GAN, training data includes only training information and does not include training information. Therefore, for example, when training is performed using a biometric likelihood estimation model as a classifier and a biometric information generation model as a generator, the trained generator often generates biometric information similar to the training information prepared as training data.

[0161] For example, the predetermined type of biometric information is an iris image, and the biometric information class to which the iris image belongs is an iris image class. When trained using general adversarial learning, the trained generator (biometric information generation model) often generates iris images that are similar to the iris images of an individual prepared as training data from among iris images belonging to the iris image class (i.e., iris images of the entire person).

[0162] In this embodiment, as described above, the likelihood that a biometric image input into the biometric likelihood estimation model belongs to a biometric information class among multiple classes including not only the biometric information class but also other classes is determined as a biometric likelihood. Then, using this biometric likelihood, the biometric information generation model can be trained to generate biometric information that resembles a predetermined type of biometric information. This allows the biometric information generation model to be trained to generate biometric information within a wider range of biometric information classes than is possible using general adversarial learning. Therefore, it is more likely that simulated biometric information (simulated biometric information of a new individual) that can be identified as an individual different from the individual included in the training biometric information can be generated.

[0163] (Operations and Effects) As described above, according to this embodiment, the biometric likelihood indicates the likelihood that biometric information belongs to a biometric information class among a plurality of classes including a biometric information class corresponding to a predetermined type of biometric information.

[0164] This increases the possibility of generating a predetermined type of biometric information that identifies an individual different from the individual included in the training biometric information, thereby making it possible to generate simulated biometric information of a new individual that resembles the predetermined type of biometric information.

[0165] Fourth Embodiment In this embodiment, a method for training a biometric information generation model to generate higher quality biometric information will be described. Note that descriptions that overlap with other embodiments will be omitted as appropriate for the sake of brevity.

[0166] As shown in FIG. 14, the information processing device 400 according to this embodiment includes a first acquisition unit 110 and a second acquisition unit 120 similar to those in the first embodiment, and a learning unit 430 .

[0167] The learning unit 430 learns the biometric information generation model.

[0168] The learning unit 430 includes a local evaluation unit 431 and an update unit 432 .

[0169] The local evaluation unit 431 calculates a local quality score indicating the quality of one or more pieces of local information obtained by dividing at least a part of the simulated biometric information.

[0170] The update unit 432 further uses the one or more local quality scores to update one or more model parameters included in the bioinformation generation model.

[0171] The information processing device 400 executes information processing as shown in FIG.

[0172] Steps S110 and S120 are executed in the same manner as in the first embodiment.

[0173] The learning unit 430 learns the biometric information generation model (step S430).

[0174] More specifically, for example, the learning unit 430 executes the following learning process (step S430).

[0175] The local evaluation unit 431 calculates a local quality score indicating the quality of one or more pieces of local information obtained by dividing at least a part of the simulated biometric information (step S431).

[0176] The update unit 432 further uses the one or more local quality scores to update one or more model parameters included in the biometric information generation model (step S432).

[0177] (Regarding the Local Evaluation Unit 431) As shown in FIG. 16, for example, the local evaluation unit 431 includes a division unit 431a and a local quality score calculation unit 431b.

[0178] The dividing unit 431a divides the entire simulated biometric information or partial information including an occluded portion into a plurality of local information.

[0179] The local quality score calculation unit 431b calculates a local quality score for each of the multiple pieces of local information based on the size of the occluded portion.

[0180] The local evaluation unit 431 executes a local evaluation process (step S431) as shown in FIG. 17, for example.

[0181] The dividing unit 431a divides the entire simulated biometric information or partial information including the occluded portion into a plurality of local information (step S431a).

[0182] The local quality score calculation unit 431b calculates a local quality score for each of the multiple pieces of local information based on the size of the occluded portion (step S431b).

[0183] (Regarding the Dividing Unit 431a) The dividing unit 431a divides the simulated biometric information into a plurality of pieces of local information.

[0184] The target to be divided by the dividing unit 431a may be, for example, the entire simulated biometric information or partial information. The partial information is a part of the simulated biometric information, and includes, for example, an occluded portion where occlusion occurs.

[0185] Occlusion occurs when, for example, a specific part included in the simulated biometric information is partially or entirely hidden by another part or object of the biometric information. For example, if the simulated biometric information is a simulated iris image that simulates an iris image, occlusion occurs when an eyelid, an intrairis reflection, a spectacle reflection, or the like covers part or all of the anterior part of the iris. Such occlusion often occurs locally in a part of the iris.

[0186] At least one of the shape and size of the partial information may be predetermined, or may be set according to the size of the obstructed portion. The size may include, for example, at least one of the length in two intersecting directions (for example, directions perpendicular to each other, such as the vertical and horizontal directions), the area, etc. The partial information may have, for example, a predetermined shape and may be a portion that is in contact with the outer edge of the obstructed portion, or may be a portion that is a predetermined length larger than the outer edge of the obstructed portion.

[0187] The number of divisions may be determined, for example, by the number of divisions in two intersecting directions, such as vertically and horizontally. Furthermore, at least one of the shape and size of the local information may be determined in advance, or may be set according to the size of the occluded portion. The size of the local information may be determined by the length of the two intersecting directions.

[0188] (Regarding the local quality score calculation unit 431b) The local quality score calculation unit 431b calculates a local quality score indicating the quality of each of the multiple pieces of local information. The local quality score may include, for example, a score corresponding to each of one or more quality items of the simulated biometric information.

[0189] The quality items are items that affect the quality of the simulated biometric information. For example, if the simulated biometric information is a simulated iris image that imitates an iris image, the quality items may include at least one of an occlusion rate, a blur rate, a circularity rate, etc.

[0190] The occlusion ratio is the ratio of the area of ​​the occluded portion to the area of ​​the local information. The area is expressed, for example, by the number of pixels, but is not limited to this. The blur ratio is a value that represents the degree of blurring of the image. The circularity ratio is a value that represents the degree to which the iris is close to being circular, and may be, for example, the difference in radii between the maximum and minimum circles that enclose the iris, or the ratio between the minor axis and the major axis when the iris is an ellipse.

[0191] The occlusion ratio is an example of a local quality score based on the size of the occluded portion. For example, the local quality score calculation unit 431b may calculate the local quality score in accordance with a predetermined criterion using the size of the occluded portion. The predetermined criterion is, for example, a formula, a table, or the like that increases the local quality score as the size of the occluded portion in the local information increases. In more detail, for example, when the occlusion ratio is used, the predetermined criterion may be a formula that calculates the ratio of the area of ​​the occluded portion to the area of ​​the local information. Note that the predetermined criterion is not limited to the one exemplified here.

[0192] (Regarding the Update Unit 432) The update unit 432 updates one or more model parameters included in the biometric information generation model using, for example, the simulated biological likelihood and one or more partial quality scores.

[0193] As shown in FIG. 18, for example, the update unit 432 includes a first loss calculation unit 232a, a second loss calculation unit 432b, a third loss calculation unit 432c, and a parameter update unit 432d, which are similar to those in the second embodiment.

[0194] The second loss calculation unit 432b uses each of the one or more local quality scores to calculate a second loss, which is a loss related to each of the one or more local quality scores.

[0195] The third loss calculation unit 432c calculates a third loss by weighting the first loss and one or more second losses using a first weight for the first loss and one or more second weights for one or more second losses.

[0196] The parameter update unit 432d updates one or more model parameters using the third loss.

[0197] The update unit 432 executes an update process (step S432) as shown in FIG. 19, for example.

[0198] Step S232a is executed in the same manner as in the second embodiment.

[0199] The second loss calculation unit 432b uses each of the one or more local quality scores to calculate a second loss, which is a loss related to each of the one or more local quality scores (step S432b).

[0200] The third loss calculation unit 432c calculates a third loss by weighting the first loss and one or more second losses using a first weight for the first loss and one or more second weights for one or more second losses (step S432c).

[0201] The parameter update unit 432d updates one or more model parameters using the third loss (step S432d).

[0202] (Second Loss Calculation Unit 432b) The second loss calculation unit 432b calculates a mean square error as the second loss, for example, using each of one or more local quality scores.

[0203] In more detail, for example, assume that the predetermined type of biometric information is an iris image, and the simulated biometric information is a simulated iris image that imitates an iris image. Examples of the second loss in this case include the following examples 1 and 2. The second loss examples 1 and 2 are examples of the second loss that differ depending on whether the local quality score has one component (a scalar quantity) or multiple components (a vector quantity).

[0204] (Example 1 of Second Loss) A simulated iris image x generated by a biometric information generation model ｉ The local quality score C for the jth local region in ｉ ｊ is a scalar quantity. Also, the local quality score C ｉ ｊ is normalized to 0 or more and 1 or less, and the closer it is to 1, the higher the quality. Here, i is an integer between 1 and B or less. B represents, for example, the batch size in mini-batch learning. j is an integer between 1 and M or less. Here, M is the number of pieces of local information.

[0205] In this case, the second loss may be a loss expressed by the following equation (4) using, for example, a mean square error.

[0206]

[0207] (Example 2 of Second Loss) A simulated iris image x generated by a biometric information generation model ｉ The local quality score for the j-th local region in is a vector, and the value of the k-th component when k is an integer between 1 and N is C ｉ ｊ，ｋ Here, N is the number of components included in the local quality score, for example, the number of quality items.

[0208] Also, C ｉ ｊ，ｋ is normalized to 0 or more and 1 or less, and the closer it is to 1, the higher the quality. Here, i is an integer between 1 and B or less. B represents, for example, the batch size in mini-batch learning. j is an integer between 1 and M or less. Here, M is the number of pieces of local information.

[0209] In this case, the second loss may be a loss expressed by the following equation (5) using, for example, a mean square error.

[0210]

[0211] The second loss is not limited to the above example, and may be, for example, a mean absolute error or the like.

[0212] (Regarding the third loss calculation unit 432c) The third loss calculation unit 432c calculates a third loss in which the first loss and one or more second losses are weighted, using a first weight for the first loss and one or more second weights for the one or more second losses.

[0213] The third loss is calculated using a first weight related to the first loss and one or more second weights related to one or more second losses.

[0214] When there are multiple second losses, one second weight may be used for the multiple second losses. Multiple second weights may be used for each of the multiple second losses. The multiple second losses may be grouped into multiple second loss groups, and multiple second weights may be used for each of the multiple second loss groups. Each second loss group includes at least one second loss. The grouping method may be determined in advance, for example.

[0215] The first weight may be, for example, a predetermined fixed value, or may be changed according to the progress of learning of the biological information generation model.

[0216] In detail, for example, when the first weight is changed, the first weight may be decreased according to the progress of learning of the biological information generation model.

[0217] Here, the first loss is a loss related to a simulated biometric likelihood that represents the likelihood of the biometric information being of a predetermined type. By decreasing the first weight according to the progress of learning of the biometric information generation model, learning can be performed in the early stage of learning with an emphasis on generating simulated biometric information that resembles the predetermined type of biometric information. In this case, the initial value of the first weight may be determined in advance.

[0218] The one or more second weights may be, for example, predetermined fixed values, or may be changed according to the progress of learning of the biological information generation model.

[0219] In detail, for example, when one or more second weights are changed, the one or more second weights may be increased according to the progress of learning of the biological information generation model, for example.

[0220] Here, the second loss is a loss related to the local quality score. By increasing the second weight according to the progress of learning of the biometric information generation model, it is possible to perform learning that emphasizes the generation of high-quality simulated biometric information with a high local quality score as learning progresses. In this case, the initial values ​​of one or more second weights may be determined in advance.

[0221] A combination of decreasing the first weight and increasing one or more second weights may be used depending on the progress of learning of the biometric information generation model. This allows the emphasis in the early stages of learning of the biometric information generation model to be placed on generating simulated biometric information that resembles a predetermined type of biometric information, and as learning progresses, the emphasis can be placed on generating high-quality simulated biometric information. This makes it possible to train the biometric image generation model so as to prevent divergence of losses and stably generate high-quality biometric information of a predetermined type, compared to when the first weight and the second weight are constant.

[0222] The method for setting the first loss and one or more second losses is not limited to the example given here.

[0223] (Regarding the parameter update unit 432d) The parameter update unit 432d updates one or more model parameters included in the biometric information generation model, for example, to reduce the third loss. In detail, for example, the parameter update unit 432d updates one or more model parameters included in the biometric information generation model using a general optimization method such as a search method such as a gradient method or a grid search. Note that the method of updating the model parameters is not limited to the example given here.

[0224] (Actions and Effects) As described above, according to this embodiment, the learning unit 430 includes a local evaluation unit 431 and an update unit 432. The local evaluation unit 431 calculates a local quality score indicating the quality of each of one or more pieces of local information obtained by dividing at least a portion of the simulated biological information. The update unit 432 further uses the one or more local quality scores to update one or more model parameters included in the biological information generation model.

[0225] This allows the biometric information generation model to be trained using the local quality score that indicates the quality of the local information so that high-quality simulated biometric information is generated. Therefore, high-quality simulated biometric information can be generated.

[0226] According to this embodiment, the update unit 432 includes a first loss calculation unit 232a, a second loss calculation unit 432b, a third loss calculation unit 432c, and a parameter update unit 432d.

[0227] The first loss calculation unit 232a calculates a first loss using the simulated biological likelihood. The second loss calculation unit 432b uses one or more local quality scores to calculate a second loss, which is a loss related to each of the one or more local quality scores. The third loss calculation unit 432c uses a first weight related to the first loss and one or more second weights related to the one or more second losses to calculate a third loss by weighting the first loss and one or more second losses. The parameter update unit 432d uses the third loss to update one or more model parameters.

[0228] This allows for the generation of simulated biometric information that resembles a predetermined type of biometric information. Furthermore, using the local quality score indicating the quality of the local information, the biometric information generation model can be trained to generate high-quality simulated biometric information. Therefore, it becomes possible to generate high-quality simulated biometric information that resembles a predetermined type of biometric information.

[0229] According to this embodiment, the local evaluation unit 431 includes a division unit 431a and a local quality score calculation unit 431b. The division unit 431a divides the entire simulated biometric information or partial information including an occluded portion into multiple pieces of local information. The local quality score calculation unit 431b calculates a local quality score for each of the multiple pieces of local information based on the size of the occluded portion.

[0230] This makes it possible to generate simulated biometric information that is unlikely to include occluded areas, thereby enabling the generation of high-quality simulated biometric information.

[0231] According to this embodiment, at least one of the first weight and one or more second weights is changed according to the progress of learning of the biological information generation model.

[0232] This allows the relative emphasis to be placed on generating simulated biometric information that resembles a predetermined type of biometric information or generating high-quality simulated biometric information, depending on the progress of the learning of the biometric information generation model, thereby enabling the generation of high-quality simulated biometric information that resembles a predetermined type of biometric information and stable convergence of the learning of the biometric information generation model.

[0233] According to this embodiment, when the first weight is changed, the first weight is decreased according to the progress of the learning of the biological information generation model, and when one or more second weights are changed, the one or more second weights are increased according to the progress of the learning of the biological information generation model.

[0234] This allows the emphasis to be placed on generating simulated biometric information that resembles a predetermined type of biometric information in the early stages of learning the biometric information generation model, and on generating high-quality simulated biometric information as the learning progresses, thereby enabling the generation of high-quality simulated biometric information that resembles a predetermined type of biometric information and stable convergence of the learning of the biometric information generation model.

[0235] [Embodiment 5] In this embodiment, an example of a method for generating new individual biometric information will be described. The new individual biometric information is simulated biometric information that identifies an individual different from an individual included in predetermined and prepared biometric information such as training biometric information. Note that descriptions that overlap with other embodiments will be omitted as appropriate for the sake of brevity.

[0236] The information processing device 500 according to this embodiment includes a first acquisition unit 110 and a second acquisition unit 120 similar to those in the first embodiment, an overall quality evaluation unit 550, and a first discrimination unit 560, as shown in FIG. 20, for example.

[0237] The overall quality evaluation unit 550 calculates an overall quality score that indicates the overall quality of the simulated biometric information.

[0238] The first determination unit 560 determines whether or not the simulated biometric information belongs to a predetermined target class by using the second simulated feature based on the simulated biometric information.

[0239] The information processing device 500 executes information processing such as that shown in FIG. 21, for example.

[0240] Steps S110 and S120 are executed in the same manner as in the first embodiment.

[0241] The overall quality evaluation unit 550 calculates an overall quality score indicating the overall quality of the simulated biometric information (step S550).

[0242] The first determination unit 560 determines whether the simulated biometric information belongs to a predetermined target class using the second simulated feature based on the simulated biometric information (step S560).

[0243] (Overall Quality Evaluation Unit 550) The overall quality evaluation unit 550 calculates, for example, an overall quality score indicating the overall quality of the simulated biometric information.

[0244] The overall quality score may include a value corresponding to at least one element, such as the degree of blur of a specific part included in the biometric information, the size (e.g., area) of the specific part, or the likelihood of a living body. The overall quality score may include one or more values ​​corresponding to each element, or may include one or more values ​​integrating multiple elements. The overall quality score may be expressed as a vector or a scalar quantity.

[0245] The overall quality assessment unit 550 may use, for example, second simulated features extracted from the simulated biometric information to calculate the overall quality score. Generally, the accuracy of biometric authentication tends to improve as the norms (e.g., L1 norm, L2 norm) of the features extracted from the biometric information using the extraction model increase.

[0246] Therefore, the overall quality evaluation unit 550 may extract the second simulated feature using, for example, an extraction model that outputs a feature when biometric information is input. For example, the overall quality evaluation unit 550 may input simulated biometric information to the extraction model to acquire the second simulated feature. Then, the overall quality evaluation unit 550 may calculate the norm of the second simulated feature as the overall quality score.

[0247] The method for calculating the overall quality score is not limited to the example described here.

[0248] (Regarding the first discrimination unit 560) The first discrimination unit 560 may, for example, determine whether the simulated biometric information belongs to a target class by using the comparison result between a second simulated feature based on the simulated biometric information and a group of class features of a predetermined target class.

[0249] The object class corresponds to, for example, a biometric characteristic of an individual. For example, if the predetermined and prepared biometric information is derived from the right eye of a single individual, the biometric information belongs to a common object class. For example, if the predetermined and prepared biometric information is derived from the right eyes of multiple individuals, the biometric information belongs to multiple object classes corresponding to each of the multiple individuals. The class feature group is composed of multiple features belonging to the object class. The class feature group may be extracted in advance from the predetermined and prepared biometric information, for example, using an extraction model.

[0250] The determination process of the first determination unit 560 may be performed, for example, when the overall quality score indicating the overall quality of the biometric information indicates higher quality than a predetermined quality. In detail, for example, the first determination unit 560 may compare the overall quality score with a predetermined quality threshold, and perform the determination process when the overall quality score is equal to or higher than the quality threshold.

[0251] As shown in FIG. 22 , the first determination unit 560 includes a third acquisition unit 561 , an attribute likelihood calculation unit 562 , and an attribute determination unit 563 .

[0252] The third acquiring unit 561 acquires a second simulated feature based on the simulated biometric information.

[0253] The attribute likelihood calculation unit 562 calculates an attribute likelihood indicating the likelihood that the second simulated feature belongs to the target class, using the distribution of similarities between the class feature group and class standard features indicating the standard of the class feature group, and the similarity between the second simulated feature and the class standard.

[0254] The attribute determining unit 563 determines whether or not the simulated biometric information belongs to a target class using the attribute likelihood.

[0255] The first determination unit 560 executes a determination process (step S560) as shown in FIG. 23, for example.

[0256] The third acquiring unit 561 acquires a second simulated feature based on the simulated biometric information (step S561).

[0257] The attribute likelihood calculation unit 562 calculates an attribute likelihood indicating the likelihood that the second simulated feature belongs to the target class, using the distribution of similarities between the class feature group and class standard features that indicate the standards for the class feature group, and the similarities between the second simulated feature and the class standard features (step S562).

[0258] The attribute determining unit 563 determines whether or not the simulated biometric information belongs to the target class using the attribute likelihood (step S563).

[0259] (Regarding the third acquisition unit 561) The third acquisition unit 561 may, for example, input simulated biometric information into an extraction model and acquire second simulated features based on the simulated biometric information. When the overall quality evaluation unit 550 acquires the second simulated features, the third acquisition unit 561 may acquire the second simulated features from the overall quality evaluation unit 550.

[0260] (Regarding the Attribute Likelihood Calculation Unit 562) The attribute likelihood calculation unit 562 stores, in advance, for example, class standard features and a distribution of similarities.

[0261] The class standard feature is a feature indicating a standard of a group of class features, and is, for example, a value obtained by statistically processing a plurality of feature values ​​belonging to a target class. More specifically, the class standard feature is, for example, an average of a plurality of feature values ​​belonging to a target class.

[0262] The similarity distribution is a distribution of similarities between each of a plurality of features belonging to a target class and the class standard features. Examples of similarities include, but are not limited to, the norm of the difference between the features and cosine similarity.

[0263] The attribute likelihood calculation unit 562 calculates, for example, the similarity between the second simulated feature and the class standard feature. Then, the attribute likelihood calculation unit 562 calculates, for example, the probability that the calculated similarity will occur in a pre-stored similarity distribution. Generally, when features are extracted from the biometric information of a specific individual using an extraction model, similar features are extracted that are robust to the quality of the biometric information. Therefore, the calculated probability is an attribute likelihood that indicates the likelihood that the second simulated feature belongs to the target class.

[0264] The method for calculating the attribute likelihood is not limited to the example given here.

[0265] (Regarding the attribute determination unit 563) The attribute determination unit 563 determines whether or not the simulated biometric information belongs to a target class by using, for example, predetermined belonging conditions and attribute likelihoods. The belonging conditions are conditions for determining whether or not the simulated biometric information belongs to a target class, and include, for example, predetermined attribute thresholds.

[0266] In detail, for example, the belonging condition may be that the attribute likelihood is equal to or greater than an attribute threshold. The attribute determining unit 563 may determine that the simulated biometric information belongs to the target class when the attribute likelihood is equal to or greater than the attribute threshold. The attribute determining unit 563 may determine that the simulated biometric information does not belong to the target class when the attribute likelihood is less than the attribute threshold.

[0267] The belonging condition is not limited to the example given here, and may be, for example, a threshold value.

[0268] The attribute discrimination unit 563 may output, for example, at least one of the discrimination result and the simulated biometric information that was the subject of discrimination. The discrimination result output by the attribute discrimination unit 563 may be information that associates the simulated biometric information that was the subject of discrimination with the discrimination result related thereto.

[0269] In detail, for example, the attribute determining unit 563 may output simulated biometric information determined not to belong to the target class as biometric information of a new individual.

[0270] (Operations and Effects) As described above, according to this embodiment, the information processing device 500 includes the first discrimination unit 560. The first discrimination unit 560 uses the second simulated feature based on the simulated biometric information to determine whether or not the simulated biometric information belongs to a predetermined target class.

[0271] This allows simulated biometric information that does not belong to the target class to be selected as biometric information for a new individual, thereby enabling the generation of simulated biometric information for a new individual.

[0272] According to this embodiment, the determination of whether simulated biometric information belongs to the target class is made when the overall quality score indicating the overall quality of the simulated biometric information indicates that the quality is higher than a predetermined quality.

[0273] This allows simulated biometric information that does not belong to the target class to be selected from the high-quality simulated biometric information as biometric information for a new individual, thereby enabling the generation of high-quality simulated biometric information for a new individual.

[0274] [Embodiment 6] In this embodiment, another example of a method for generating new individual biometric information will be described. In addition, in this embodiment, another example of a method for training a biometric information generation model will also be described. Note that descriptions that overlap with other embodiments will be omitted as appropriate for the sake of brevity.

[0275] As shown in FIG. 24, the information processing device 600 according to this embodiment includes a first acquisition unit 110 similar to that of the first embodiment, a second determination unit 670, a second acquisition unit 620, and a learning unit 630.

[0276] The second discrimination unit 670 discriminates whether or not the first simulated feature belongs to a predetermined target class.

[0277] The second acquiring unit 620 acquires simulated biometric information based on the first simulated feature determined not to belong to the target class, using the biometric information generation model.

[0278] The learning unit 630 learns the biometric information generation model.

[0279] In detail, for example, the learning unit 630 includes a fourth acquisition unit 631 and an update unit 632 .

[0280] The fourth acquiring unit 631 acquires training features based on training biometric information and second simulated features based on simulated biometric information.

[0281] The update unit 632 updates one or more model parameters included in the biometric information generation model using the fourth loss related to the training biometric information and the simulated biometric information, and the fifth loss related to the training feature and the second simulated feature.

[0282] The information processing device 600 executes information processing as shown in FIG.

[0283] Step S110 is executed in the same manner as in the first embodiment.

[0284] The second discrimination unit 670 discriminates whether or not the first simulated feature belongs to a predetermined target class (step S670).

[0285] The second acquiring unit 620 acquires simulated biometric information based on the first simulated feature determined not to belong to the target class, using the biometric information generation model (step S620).

[0286] The learning unit 630 learns the biometric information generation model (step S630).

[0287] The fourth acquiring unit 631 acquires training features based on the training biometric information and second simulated features based on the simulated biometric information (step S631).

[0288] The update unit 632 updates one or more model parameters included in the biometric information generation model using the fourth loss related to the training biometric information and the simulated biometric information and the fifth loss related to the training feature and the second simulated feature (step S632).

[0289] (Regarding the second discrimination unit 670) The second discrimination unit 670 determines whether the first simulated feature belongs to the target class, for example, by using a comparison result between the first simulated feature and a predetermined group of class features.

[0290] In detail, for example, the second discrimination unit 670 may perform processing using the first simulated feature instead of the second simulated feature used by the above-described attribute likelihood calculation unit 562. That is, the second discrimination unit 670 may calculate an attribute likelihood indicating the likelihood that the first simulated feature belongs to the target class by using, for example, a distribution of similarities between a group of class features and class standard features indicating standards for the group of class features, and the similarity between the first simulated feature and the class standard.

[0291] Then, the second discrimination unit 670 may determine whether or not the first simulated feature belongs to the target class using, for example, the attribute likelihood. The determination of whether or not the first simulated feature belongs to the target class may be performed using a predetermined belonging condition and the attribute likelihood, similar to the attribute discrimination unit 563. Therefore, for example, the second discrimination unit 670 may determine that the first simulated feature does not belong to the target class when the attribute likelihood is less than an attribute threshold.

[0292] The first simulated feature may be represented by a tensor having at least one of a dimension and a norm of 1, for example.

[0293] (Second Acquisition Unit 620) The second acquisition unit 620 acquires simulated biometric information based on the first simulated feature determined not to belong to the target class, using the biometric information generation model.

[0294] (Regarding the fourth acquisition unit 631) The fourth acquisition unit 631 may acquire training features based on the training biometric information and second simulated features based on the simulated biometric information, for example, by using an extraction model that outputs features when biometric information is input. In detail, for example, the fourth acquisition unit 631 may input each of the training biometric information and the simulated biometric information to the extraction model, and acquire training features based on the training biometric information and second simulated features based on the simulated biometric information.

[0295] Note that the training feature based on the training biometric information may be prepared in advance together with the training biometric information, for example.

[0296] (Regarding the Updater 632) As shown in FIG. 26, for example, the updater 632 includes a fourth loss calculator 632a, a fifth loss calculator 632b, a sixth loss calculator 632c, and a parameter updater 632d.

[0297] The fourth loss calculation unit 632a calculates the fourth loss relating to the training biological information and the simulated biological information.

[0298] The fifth loss calculation unit 632b calculates a fifth loss related to the training features and the second simulated features.

[0299] The sixth loss calculation unit 632c calculates a sixth loss by combining the fourth loss and the fifth loss.

[0300] The parameter update unit 632d uses the sixth loss to update one or more model parameters included in the biological information generation model.

[0301] The update unit 632 executes an update process (step S632) as shown in FIG. 27, for example.

[0302] The fourth loss calculation unit 632a calculates the fourth loss related to the learning biological information and the simulated biological information (Step S632a).

[0303] The fifth loss calculating unit 632b calculates the fifth loss for the training features and the second simulated features (Step S632b).

[0304] The sixth loss calculation unit 632c calculates a sixth loss by combining the fourth loss and the fifth loss (step S632c).

[0305] The parameter update unit 632d uses the sixth loss to update one or more model parameters included in the biometric information generation model (step S632d).

[0306] (Regarding the fourth loss calculation unit 632a) The fourth loss calculation unit 632a calculates the fourth loss based on, for example, the difference between the training biometric information and the simulated biometric information. For example, in the case of training biometric information, simulated biometric information, and an image, the difference may be the absolute mean, root mean square, or the like of the difference between corresponding pixel values ​​therebetween. The root mean square is the average of squared values.

[0307] (Fifth loss calculation unit 632b) The fifth loss calculation unit 632b calculates the fifth loss based on the difference between the training features and the second simulated features. For example, when the training features and the second simulated features are represented by tensors, the difference may be the absolute mean, the root mean square, or the like of the difference between them.

[0308] (Regarding the Sixth Loss Calculation Unit 632c) The sixth loss calculation unit 632c may, for example, perform predetermined statistical processing on the fourth loss and the fifth loss to integrate the fourth loss and the fifth loss, and may use the result as the sixth loss. The statistical processing may, for example, calculate the average value of the fourth loss and the fifth loss, calculate a weighted sum in which the fourth loss and the fifth loss are weighted using predetermined fourth weights and fifth weights, or identify the maximum value of the fourth loss and the fifth loss. Note that the method of integrating the fourth loss and the fifth loss into the sixth loss is not limited to the example given here.

[0309] (Regarding the parameter update unit 632d) The parameter update unit 632d updates one or more model parameters included in the biometric information generation model, for example, to reduce the sixth loss. In detail, for example, the parameter update unit 632d updates one or more model parameters included in the biometric information generation model using a general optimization method such as a search method such as a gradient method or a grid search. Note that the method of updating the model parameters is not limited to the example given here.

[0310] According to the present embodiment, the information processing device 600 includes the second determination unit 670 that determines whether the first simulated feature belongs to a predetermined target class. The second acquisition unit 620 uses the biometric information generation model to acquire simulated biometric information based on the first simulated feature determined not to belong to the target class.

[0311] This allows simulated biometric information based on the first simulated feature to be acquired that has a high probability of being identified as a new individual, thereby enabling efficient generation of simulated biometric information of new individuals.

[0312] According to this embodiment, the information processing device 600 includes a learning unit 630 that learns the biological information generation model. The learning unit 630 includes a fourth acquisition unit 631 and an update unit 632.

[0313] The fourth acquisition unit 631 acquires training features based on the training biometric information and second simulated features based on the simulated biometric information. The update unit 632 updates one or more model parameters included in the biometric information generation model using a fourth loss related to the training biometric information and the simulated biometric information and a fifth loss related to the training features and the second simulated features.

[0314] This allows us to build a biometric information generation model that has learned the process of mapping biometric information from features, as opposed to the process of extracting features from biometric information. Therefore, it is possible to generate simulated biometric information with various variations belonging to many individuals using random numbers that follow a predetermined distribution.

[0315] In addition, by using a biometric information generation model that has learned the process of mapping biometric information from features to generate simulated biometric information from a first simulated feature that has the potential to be identified as a new individual, it becomes possible to generate simulated biometric information of a new individual.

[0316] According to this embodiment, the first simulated feature is represented by a tensor having at least one of a dimension and a norm of 1, for example.

[0317] This reduces the processing load of using the first simulated feature, for example, the processing load of determining whether the first simulated feature belongs to a target class, thereby making it possible to easily generate simulated biometric information with various variations belonging to many individuals.

[0318] Seventh Embodiment In this embodiment, an example of a method for learning an extraction model will be described. Note that descriptions that overlap with other embodiments will be omitted as appropriate for the sake of brevity.

[0319] As shown in FIG. 28 , the information processing device 700 according to this embodiment includes a first acquisition unit 110 and a second acquisition unit 120 similar to those in the first embodiment, an extraction unit 780, a second discrimination unit 770, and an extraction model update unit 790.

[0320] The extraction unit 780 acquires a second simulated feature based on the simulated biometric information using an extraction model that extracts a feature from the biometric information.

[0321] The second determination unit 770 determines whether the second simulated feature belongs to a predetermined target class.

[0322] The extraction model update unit 790 updates one or more extraction model parameters included in the extraction model.

[0323] The information processing device 700 executes information processing as shown in FIG.

[0324] Steps S110 and S120 are executed in the same manner as in the first embodiment.

[0325] The extracting unit 780 acquires a second simulated feature based on the simulated biometric information using an extraction model that extracts a feature from the biometric information (step S780).

[0326] The second discrimination unit 770 discriminates whether the second simulated feature belongs to a predetermined target class (step S770).

[0327] The extraction model update unit 790 updates one or more extraction model parameters included in the extraction model (step S790).

[0328] (Extraction Unit 780) The extraction unit 780, for example, inputs simulated biometric information into an extraction model and acquires second simulated features based on the simulated biometric information.

[0329] (Second Discrimination Section 770) The second discrimination section 770 may use, for example, a second simulated feature instead of the first simulated feature used by the second discrimination section 670.

[0330] That is, for example, the second discrimination unit 770 may determine whether the second simulated feature belongs to the target class by using, for example, a comparison result between the second simulated feature and a predetermined class feature group. The second discrimination unit 770 may calculate an attribute likelihood indicating the likelihood that the second simulated feature belongs to the target class by using, for example, a distribution of similarities between the class feature group and class standard features indicating standards for the class feature group, and a similarity between the second simulated feature and the class standard.

[0331] Then, the second discrimination unit 670 may determine whether or not the second simulated feature belongs to the target class using, for example, the attribute likelihood. The determination of whether or not the second simulated feature belongs to the target class may be performed using a predetermined belonging condition and the attribute likelihood, similar to the attribute discrimination unit 563. Therefore, for example, the second discrimination unit 770 may determine that the second simulated feature does not belong to the target class when the attribute likelihood is less than an attribute threshold.

[0332] (Extraction Model Update Unit 790) The extraction model update unit 790 may update one or more extraction model parameters included in the extraction model by using, for example, previously prepared training data. The training data includes training biometric information and a correct label indicating a class to which the training biometric information belongs.

[0333] For example, when it is determined that the second simulated feature does not belong to the target class, the extraction model update unit 790 may assign a new correct label to the second simulated feature. The new correct label is a new class different from the predetermined class group. Then, for example, the extraction model update unit 790 may add the second simulated feature determined not to belong to the target class and the new correct label to the training data.

[0334] The extraction model update unit 790 may update one or more extraction model parameters so that the second simulated feature is classified into the correct label.

[0335] In detail, for example, the extraction model updating unit 790 may estimate a class to which the second simulated feature belongs based on the similarity (e.g., cosine similarity) between the second simulated feature, training features based on training biometric information included in the training data, and the second simulated feature. The extraction model updating unit 790 may calculate a loss (e.g., mean absolute error, mean squared error, cross entropy error, etc.) based on the estimated class and the correct label. The extraction model updating unit 790 may update one or more extraction model parameters to reduce the loss using a general optimization method such as a search method such as a gradient method or a grid search.

[0336] Then, a second simulated feature based on the simulated biometric information may be further acquired using the extraction model with the updated extraction model parameters. The second determination unit 670 may further determine whether the second simulated feature belongs to a predetermined target class using the second simulated feature. In this manner, steps S780, S770, and S790 may be repeatedly executed using the simulated biometric information acquired in step S120. The condition for ending this repetition may be, for example, that the process has been repeated a predetermined number of times, or that the loss has become equal to or less than a threshold value.

[0337] (Operations and Effects) As described above, according to this embodiment, the information processing device 700 includes the extraction unit 780 , the second determination unit 770 , and the extraction model update unit 790 .

[0338] The extracting unit 780 acquires second simulated features based on the simulated biometric information using an extraction model that extracts features from the biometric information. The second determining unit 770 determines whether the second simulated features belong to a predetermined target class. The extraction model updating unit 790 updates one or more extraction model parameters included in the extraction model.

[0339] This allows the second simulated feature of the new individual to be used to train an extraction model. Therefore, it becomes possible to build an extraction model that extracts features with high class identification accuracy. [Embodiment 8] In this embodiment, an example of outputting various information will be described. Note that descriptions that overlap with other embodiments will be omitted as appropriate for the sake of brevity.

[0340] An information processing device 800 according to this embodiment includes a first acquisition unit 110 and a second acquisition unit 120 similar to those in the first embodiment, and an output unit 890, as shown in FIG. 30, for example.

[0341] The output unit 890 outputs various types of information.

[0342] The information processing device 800 executes information processing such as that shown in FIG.

[0343] Steps S110 and S120 are executed in the same manner as in the first embodiment.

[0344] The output unit 890 outputs various information (step S890).

[0345] (Regarding the Output Unit 890) The output unit 890 may output output information in which a label is associated with at least one of the first simulated feature and the simulated biometric information, as shown in, for example, Figures 3 and 4. The label is information indicating a class (e.g., individual) to which at least one of the first simulated feature and the simulated biometric information belongs.

[0346] The output unit 890 may be provided in the information processing device according to each embodiment.

[0347] For example, the output unit 890 may output the determination result of the first determination unit 560. For example, the output unit 890 may output, together with the determination result of the first determination unit 560, at least one of the second simulated feature used in the determination, the simulated biometric information from which the second simulated feature was extracted, and the first simulated feature used to generate the simulated biometric information.

[0348] For example, the output unit 890 may output the determination result of the second determination unit 670. For example, the output unit 890 may output, together with the determination result of the second determination unit 670, at least one of the first simulated feature used in the determination, the simulated biometric information based on the first simulated feature, and the second simulated feature based on the simulated biometric information.

[0349] The output may be, for example, a display or transmission to another device.

[0350] (Operations and Effects) As described above, according to this embodiment, the information processing device 800 includes the output unit 890 that outputs output information in which a label is associated with at least one of the first simulated feature and the simulated biometric information.

[0351] This makes it possible to easily recognize the correspondence between at least one of the first simulated feature and the simulated biometric information and the label.

[0352] (Variation 1) In the above-described embodiment, an example has been described in which an information processing device has each function and executes each process. The functions of the information processing device in the above-described embodiment may be provided in an information processing system composed of one or more information processing devices. In this case, the functions of the information processing device in the above-described embodiment may be shared among one or more information processing devices and provided as a whole by one or more information processing devices.

[0353] 32 is a diagram showing an example configuration of an information processing system SYS. The information processing system SYS includes, for example, a first acquisition unit 110 and a second acquisition unit 120. The first acquisition unit 110 acquires first simulated features based on random numbers according to a predetermined distribution. The second acquisition unit 120 acquires simulated biometric information based on the first simulated features using a biometric information generation model that generates simulated biometric information that simulates a predetermined type of biometric information from the first simulated features.

[0354] This information processing system SYS can generate simulated biometric information using random numbers that follow a predetermined distribution, making it possible to generate simulated biometric information with a wide variety of variations that belong to many individuals.

[0355] Although the present disclosure has been described above with reference to the embodiments, the present disclosure is not limited to the above-described embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure. Furthermore, each embodiment can be combined with other embodiments as appropriate.

[0356] In addition, although the flowcharts used in the above description show a sequence of steps (processes), the order of steps executed in each embodiment is not limited to the sequence shown in the flowcharts. In each embodiment, the order of steps shown in the diagrams can be changed as long as it does not cause any problems in terms of the content.

[0357] Some or all of the above embodiments may be described as, but are not limited to, the following supplementary notes. 1. An information processing device comprising: a first acquisition means for acquiring a first simulated feature based on a random number following a predetermined distribution; and a second acquisition means for acquiring simulated biometric information based on the first simulated feature using a biometric information generation model that generates simulated biometric information simulating a predetermined type of biometric information from the first simulated feature. 2. The information processing device described in 1., further comprising learning means for training the biometric information generation model, the learning means including: biometric likelihood estimation means for estimating a simulated biometric likelihood, which is the biometric likelihood based on the simulated biometric information, using a biometric likelihood estimation model that estimates a biometric likelihood indicating the likelihood that biometric information is the predetermined type of biometric information; and update means for updating one or more model parameters included in the biometric information generation model using the simulated biometric likelihood. 3. 2. The information processing device described in 2., wherein the learning means includes: local evaluation means that calculates, for each of one or more pieces of local information obtained by dividing at least a portion of the simulated biological information, a local quality score that indicates a quality of the local information; and update means that further uses the one or more local quality scores to update one or more model parameters included in the biological information generation model. 4. The information processing device described in 3., wherein the update means includes: first loss calculation means that calculates a first loss using the simulated biological likelihood; second loss calculation means that uses each of the one or more local quality scores to calculate a second loss that is a loss related to each of the one or more local quality scores; third loss calculation means that calculates a third loss by weighting the first loss and the one or more second losses using a first weight related to the first loss and one or more second weights related to one or more of the second losses; and parameter update means that updates the one or more model parameters using the third loss.5. The information processing device described in 3. or 4., wherein the local evaluation means includes: a division means for dividing the entire simulated biometric information or partial information including an occluded portion into a plurality of pieces of local information; and a local quality score calculation means for calculating the local quality score for each of the plurality of pieces of local information based on the size of the occluded portion. 6. The information processing device described in 4., wherein at least one of the first weight and the one or more second weights is changed according to the progress of learning of the biometric information generation model. 7. The information processing device described in 4., wherein when the first weight is changed, the first weight is decreased according to the progress of learning of the biometric information generation model, and when the one or more second weights are changed, the one or more second weights are increased according to the progress of learning of the biometric information generation model. 8. The information processing device described in any one of 2. to 7., wherein the biometric likelihood indicates the likelihood that the biometric information belongs to a biometric information class among a plurality of classes including a biometric information class corresponding to the predetermined type of biometric information. 9. The information processing device according to any one of 1. to 8., further comprising first determination means for determining whether the simulated biometric information belongs to a predetermined target class using second simulated features based on the simulated biometric information. 10. The information processing device according to 9., wherein the determination of whether the simulated biometric information belongs to the target class is made when an overall quality score indicating the overall quality of the simulated biometric information indicates a higher quality than a predetermined quality. 11. The information processing device according to any one of 1. to 10., further comprising second determination means for determining whether the first simulated features belong to a predetermined target class, wherein the second acquisition means uses the biometric information generation model to acquire simulated biometric information based on the first simulated features determined not to belong to the target class.12. The information processing device described in 11., further comprising learning means for learning the biometric information generation model, wherein the learning means includes: fourth acquisition means for acquiring training features based on training biometric information and second simulated features based on the simulated biometric information; and update means for updating one or more model parameters included in the biometric information generation model using a fourth loss related to the training biometric information and the simulated biometric information and a fifth loss related to the training features and the second simulated features. 13. The information processing device described in any one of 1. to 12., wherein the first simulated features are expressed as a matrix with at least one of a dimension and a norm of 1. 14. The information processing device described in any one of 1. to 13., wherein the predetermined type of biometric information is an image showing an iris, and includes a rectangular iris image having sides corresponding to the radial and circumferential directions of the iris, and the first simulated features include features in a two-dimensional map format in which the random numbers are arranged as components of a two-dimensional tensor. 15. 16. The information processing device according to any one of 1. to 14., further comprising: extraction means for acquiring second simulated features based on the simulated biometric information using an extraction model that extracts features from the biometric information; second determination means for determining whether the second simulated features belong to a predetermined target class; and extraction model update means for updating one or more extraction model parameters included in the extraction model, wherein the extraction model update means updates the one or more extraction model parameters using previously prepared learning biometric information and a correct label for the learning biometric information, and a second simulated feature determined not to belong to the target class and a correct label assigned to the second simulated feature. 17. The information processing device according to any one of 1. to 15., further comprising output means for outputting output information in which a label is associated with at least one of the first simulated features and the simulated biometric information. An information processing system comprising: a first acquisition means for acquiring a first simulated feature based on a random number according to a predetermined distribution; and a second acquisition means for acquiring the simulated biometric information based on the first simulated feature using a biometric information generation model that generates simulated biometric information that imitates a predetermined type of biometric information from the first simulated feature.18. The information processing system described in 17., further comprising learning means for learning the biometric information generation model, wherein the learning means includes: biometric likelihood estimation means for estimating a simulated biometric likelihood, which is a biometric likelihood based on the simulated biometric information, using a biometric likelihood estimation model that estimates a biometric likelihood indicating the likelihood that the biometric information is the predetermined type of biometric information; and update means for updating one or more model parameters included in the biometric information generation model using the simulated biometric likelihood. 19. The information processing system described in 18., wherein the learning means includes: local evaluation means for calculating, for each of one or more pieces of local information obtained by dividing at least a portion of the simulated biometric information, a local quality score indicating the quality of the local information; and update means for updating one or more model parameters included in the biometric information generation model, further using the one or more local quality scores. 20. 21. The information processing system according to 19. or 20., wherein the updating means includes: first loss calculation means that calculates a first loss using the simulated biological likelihood; second loss calculation means that uses each of the one or more local quality scores to calculate a second loss which is a loss related to each of the one or more local quality scores; third loss calculation means that calculates a third loss by weighting the first loss and the one or more second losses using a first weight related to the first loss and one or more second weights related to one or more of the second losses; and parameter updating means that updates the one or more model parameters using the third loss. 22. The information processing system according to 19. or 20., wherein the local evaluation means includes: division means that divides the entire simulated biological information or partial information including an occluded portion where occlusion has occurred into a plurality of local information; and local quality score calculation means that calculates the local quality score for each of the plurality of local information based on the size of the occluded portion. 21. The information processing system according to 20., wherein at least one of the first weight and the one or more second weights is changed according to a progress of learning of the biological information generation model.23. The information processing system described in 20., wherein when the first weight is changed, the first weight is decreased according to the progress of learning of the biometric information generation model, and when the one or more second weights are changed, the one or more second weights are increased according to the progress of learning of the biometric information generation model. 24. The information processing system described in any one of 18. to 23., wherein the biometric likelihood indicates the likelihood that the biometric information belongs to the biometric information class among a plurality of classes including a biometric information class corresponding to the predetermined type of biometric information. 25. The information processing system described in any one of 17. to 24., further comprising first determination means for determining whether the simulated biometric information belongs to a predetermined target class using second simulated features based on the simulated biometric information. 26. The information processing system described in 25., wherein the determination of whether the simulated biometric information belongs to the target class is made when an overall quality score indicating the overall quality of the simulated biometric information indicates a higher quality than a predetermined quality. 27. The information processing system of any one of 17. to 26., further comprising second determination means for determining whether the first simulated feature belongs to a predetermined target class, wherein the second acquisition means uses the biometric information generation model to acquire simulated biometric information based on the first simulated feature determined not to belong to the target class. 28. The information processing system of 27., further comprising learning means for training the biometric information generation model, wherein the learning means includes: fourth acquisition means for acquiring training features based on training biometric information and second simulated features based on the simulated biometric information, and update means for updating one or more model parameters included in the biometric information generation model using a fourth loss related to the training biometric information and the simulated biometric information, and a fifth loss related to the training features and the second simulated features. 29. The information processing system of any one of 17. to 28., wherein the first simulated feature is represented by a matrix in which at least one of a dimension and a norm is 1.30. The information processing system described in any one of 17. to 29., wherein the predetermined type of biometric information is an image showing an iris and includes a rectangular iris image having sides corresponding to the radial and circumferential directions of the iris, and the first simulated feature includes a feature in a two-dimensional map format in which the random numbers are arranged as components of a two-dimensional tensor. 31. The information processing system described in any one of 17. to 30., further comprising: extraction means for acquiring second simulated features based on the simulated biometric information using an extraction model that extracts features from the biometric information; second determination means for determining whether the second simulated features belong to a predetermined target class; and extraction model update means for updating one or more extraction model parameters included in the extraction model, wherein the extraction model update means updates the one or more extraction model parameters using previously prepared training biometric information and a correct label for the training biometric information, second simulated features determined not to belong to the target class, and a correct label assigned to the second simulated features. 32. The information processing system according to any one of 17. to 31., further comprising output means for outputting output information in which a label is associated with at least one of the first simulated feature and the simulated biometric information. 33. An information processing method in which one or more computers acquire first simulated feature(s) based on random numbers according to a predetermined distribution, and acquire the simulated biometric information based on the first simulated feature(s) using a biometric information generation model that generates simulated biometric information that simulates a predetermined type of biometric information from the first simulated feature(s). 34. The information processing method according to 33., further comprising: training the biometric information generation model, estimating a simulated biometric likelihood, which is the biometric likelihood based on the simulated biometric information, using a biometric likelihood estimation model that estimates a biometric likelihood indicating the likelihood that biometric information is the predetermined type of biometric information, and updating one or more model parameters included in the biometric information generation model using the simulated biometric likelihood.35. The information processing method according to 34., wherein training the biometric information generation model includes: calculating a local quality score indicating the quality of one or more pieces of local information obtained by dividing at least a portion of the simulated biometric information, for each of the pieces of local information, and further using the one or more local quality scores to update one or more model parameters included in the biometric information generation model. 36. The information processing method according to 35., wherein updating the one or more model parameters includes: calculating a first loss using the simulated biometric likelihood; calculating a second loss that is a loss related to each of the one or more local quality scores, using each of the one or more local quality scores; calculating a third loss that weights the first loss and the one or more second losses, using a first weight related to the first loss and one or more second weights related to one or more of the second losses; and updating the one or more model parameters using the third loss. 37. The information processing method according to 35. or 36., wherein calculating the local quality score involves dividing the entire simulated biometric information or partial information including an occluded portion into a plurality of pieces of local information, and calculating the local quality score for each of the plurality of pieces of local information based on the size of the occluded portion. 38. The information processing method according to 36., wherein at least one of the first weight and the one or more second weights is changed according to the progress of learning of the biometric information generation model. 39. The information processing method according to 36., wherein when the first weight is changed, the first weight is decreased according to the progress of learning of the biometric information generation model, and when the one or more second weights are changed, the one or more second weights are increased according to the progress of learning of the biometric information generation model. 40. The information processing method according to any one of 34. to 39., wherein the biometric likelihood indicates the likelihood that the biometric information belongs to a biometric information class among a plurality of classes including a biometric information class corresponding to the predetermined type of biometric information. 41. The information processing method according to any one of items 33 to 40, further comprising determining whether or not the simulated biometric information belongs to a predetermined target class using a second simulated feature based on the simulated biometric information.42. The information processing method according to 41., wherein determining whether the simulated biometric information belongs to the target class is performed when an overall quality score indicating the overall quality of the simulated biometric information indicates a higher quality than a predetermined quality. 43. The information processing method according to any one of 33. to 42., further comprising determining whether the first simulated feature belongs to a predetermined target class, and acquiring the simulated biometric information based on the first simulated feature by using the biometric information generation model to acquire simulated biometric information based on the first simulated feature determined not to belong to the target class. 44. The information processing method according to 43., further comprising: training the biometric information generation model; acquiring training features based on training biometric information and second simulated features based on the simulated biometric information; and updating one or more model parameters included in the biometric information generation model using a fourth loss related to the training biometric information and the simulated biometric information, and a fifth loss related to the training features and the second simulated feature. 45. The information processing method described in any one of 33. to 44., wherein the first simulated feature is expressed as a matrix with at least one of a dimension and a norm of 1. 46. The information processing method described in any one of 33. to 45., wherein the predetermined type of biometric information is an image showing an iris and includes a rectangular iris image having sides corresponding to the radial and circumferential directions of the iris, and the first simulated feature includes a feature in a two-dimensional map format in which the random numbers are arranged as components of a two-dimensional tensor. 47. The information processing method described in any one of Items 33 to 46, further comprising: acquiring second simulated features based on the simulated biometric information using an extraction model that extracts features from the biometric information; determining whether the second simulated features belong to a predetermined target class; and updating one or more extraction model parameters included in the extraction model; and updating the one or more extraction model parameters includes updating the one or more extraction model parameters using previously prepared training biometric information and a correct label for the training biometric information, and second simulated features determined not to belong to the target class and a correct label assigned to the second simulated features.48. The information processing method according to any one of 33 to 47, further comprising outputting output information in which a label is associated with at least one of the first simulated feature and the simulated biometric information. 49. A program for causing one or more computers to execute the information processing method according to any one of 33 to 48. 50. A recording medium having recorded thereon a program for causing one or more computers to execute the information processing method according to any one of 33 to 48.

[0358] 100, 200, 300, 400, 500, 600, 700, 800 Information processing device 110 First acquisition unit 120, 620 Second acquisition unit 230, 330, 430, 630 Learning unit 231, 331 Biological likelihood estimation unit 232, 432, 632 Update unit 232a First loss calculation unit 232b Parameter update unit 340 Likelihood model learning unit 341 Likelihood model update unit 341a Likelihood loss calculation unit 341b Likelihood parameter update unit 342 Training data storage unit 431 Local evaluation unit 431a Segmentation unit 431b Local quality score calculation unit 432b Second loss calculation unit 432c Third loss calculation unit 432d, 632d Parameter update unit 550 Overall quality evaluation unit 560 First discrimination unit 561 Third acquisition unit 562 Attribute likelihood calculation unit 563 Attribute discrimination unit 631 Fourth acquisition unit 632a Fourth loss calculation unit 632b Fifth loss calculation unit 632c Sixth loss calculation unit 670, 770 Second discrimination unit 780 Extraction unit 790 Extraction model update unit 890 Output unit

Claims

1. An information processing device comprising: a first acquisition means for acquiring a first simulated feature based on random numbers following a predetermined distribution; and a second acquisition means for acquiring the simulated biometric information based on the first simulated feature using a biometric information generation model that generates simulated biometric information that imitates a predetermined type of biometric information from the first simulated feature.

2. An information processing device as described in claim 1, further comprising a learning means for learning the biometric information generation model, wherein the learning means includes: a biometric likelihood estimation means for estimating a simulated biometric likelihood, which is the biometric likelihood based on the simulated biometric information, using a biometric likelihood estimation model that estimates a biometric likelihood indicating the likelihood that the biometric information is the predetermined type of biometric information; and an update means for updating one or more model parameters included in the biometric information generation model using the simulated biometric likelihood.

3. The information processing device according to claim 2, wherein the learning means includes: a local evaluation means for calculating a local quality score indicating the quality of one or more pieces of local information obtained by dividing at least a portion of the simulated biometric information; and an update means for further using one or more of the local quality scores to update one or more model parameters included in the biometric information generation model.

4. The information processing device according to claim 3, wherein the updating means includes: a first loss calculation means for calculating a first loss using the simulated organism likelihood; a second loss calculation means for calculating a second loss, which is a loss related to each of the one or more local quality scores, using each of the one or more local quality scores; a third loss calculation means for calculating a third loss by weighting the first loss and the one or more second losses, using a first weight related to the first loss and one or more second weights related to one or more of the second losses; and a parameter updating means for updating the one or more model parameters using the third loss.

5. The information processing device according to claim 3 or 4, wherein the local evaluation means includes: a division means for dividing the entire simulated biometric information or partial information including an occluded portion into a plurality of pieces of local information; and a local quality score calculation means for calculating the local quality score for each of the plurality of pieces of local information based on the size of the occluded portion.

6. The information processing device according to claim 4, wherein at least one of the first weight and the one or more second weights is changed according to the progress of learning of the biological information generation model.

7. An information processing device as described in claim 4, wherein when the first weight is changed, the first weight is decreased according to the progress of learning of the biological information generation model, and when the one or more second weights are changed, the one or more second weights are increased according to the progress of learning of the biological information generation model.

8. An information processing device according to any one of claims 2 to 7, wherein the biometric likelihood indicates the likelihood that the biometric information belongs to a biometric information class among a plurality of classes including a biometric information class corresponding to the predetermined type of biometric information.

9. An information processing device according to any one of claims 1 to 8, further comprising a first determination means for determining whether the simulated biometric information belongs to a predetermined target class using a second simulated feature based on the simulated biometric information.

10. An information processing device according to claim 9, wherein the determination of whether the simulated biometric information belongs to the target class is made when an overall quality score indicating the overall quality of the simulated biometric information indicates a higher quality than a predetermined quality.

11. An information processing device as described in any one of claims 1 to 10, further comprising a second determination means for determining whether the first simulated feature belongs to a predetermined target class, and the second acquisition means uses the biometric information generation model to acquire simulated biometric information based on the first simulated feature determined not to belong to the target class.

12. An information processing device as described in claim 11, further comprising a learning means for learning the biometric information generation model, wherein the learning means includes: a fourth acquisition means for acquiring training features based on the training biometric information and second simulated features based on the simulated biometric information; and an update means for updating one or more model parameters included in the biometric information generation model using a fourth loss related to the training biometric information and the simulated biometric information, and a fifth loss related to the training features and the second simulated features.

13. The information processing device according to any one of claims 1 to 12, wherein the first simulated feature is expressed as a matrix in which at least one of a dimension and a norm is 1.

14. An information processing device according to any one of claims 1 to 13, wherein the predetermined type of biometric information is an image showing an iris, and includes a rectangular iris image having sides corresponding to the radial and circumferential directions of the iris, and the first simulated feature includes a feature in a two-dimensional map format in which the random numbers are arranged as components of a two-dimensional tensor.

15. An information processing device according to any one of claims 1 to 14, further comprising: an extraction means for acquiring second simulated features based on the simulated biometric information using an extraction model that extracts features from the biometric information; a second determination means for determining whether the second simulated features belong to a predetermined target class; and an extraction model update means for updating one or more extraction model parameters included in the extraction model, wherein the extraction model update means updates the one or more extraction model parameters using previously prepared training biometric information and a correct label for the training biometric information, a second simulated feature determined not to belong to the target class, and a correct label assigned to the second simulated feature.

16. The information processing device according to any one of claims 1 to 15, further comprising output means for outputting output information in which a label is associated with at least one of the first simulated feature and the simulated biometric information.

17. An information processing system comprising: a first acquisition means for acquiring a first simulated feature based on random numbers following a predetermined distribution; and a second acquisition means for acquiring the simulated biometric information based on the first simulated feature using a biometric information generation model that generates simulated biometric information that imitates a predetermined type of biometric information from the first simulated feature.

18. An information processing method in which one or more computers acquire first simulated features based on random numbers following a predetermined distribution, and acquire simulated biometric information based on the first simulated features using a biometric information generation model that generates simulated biometric information that imitates a predetermined type of biometric information from the first simulated features.

19. A recording medium having recorded thereon a program for causing one or more computers to acquire first simulated features based on random numbers following a predetermined distribution, and to acquire simulated biometric information based on the first simulated features using a biometric information generation model that generates simulated biometric information that imitates a predetermined type of biometric information from the first simulated features.

Images