Information processing system, information processing method and program

Abstract

An information processing system (100) is provided with an extraction unit (112), a calculation unit (113), and an updating unit (114). The extraction unit (112) uses an extraction model that performs learning with training images including an iris as input, to extract iris information related to the iris. The calculation unit (113) uses resolution of an iris region included in the training images to calculate loss between the iris information and correct answer information. The updating unit (114) uses the calculated loss to update the value of a parameter of the extraction model.

Description

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

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

[0002] Various techniques using learning models have been proposed to extract iris features (feature amounts of the iris) used in iris authentication, etc. (See, for example, Patent Documents 1 to 3.) For example, Patent Document 2 describes the use of L2 softmax loss, cosine loss, ArcFace, CosFace, SphereFace, AdaCos, or the like as a loss.

[0003] Non-Patent Documents 1 to 4 disclose examples of loss functions such as CosFace, ArcFace, MagFace, and T-center.

[0004] International Publication No. 2022 / 208606 International Publication No. 2022 / 195819 International Publication No. 2022 / 185436

[0005] Hao Wang and seven others, "CosFace: Large Margin Cosine Loss for Deep Face Recognition," [online], CVPR 2018, [Retrieved December 20, 2022], Internet <URL: https: / / openaccess.thecvf.com / content_cvpr_2018 / html / Wang_CosFace_Large_Margin_CVPR_2018_paper.html>

[0006] Jiankang Deng, et al., "ArcFace: Additive Angular Margin Loss for Deep Face Recognition", [online], CVPR 2019, [searched on December 20, 2022], Internet <URL: https: / / openaccess.thecvf.com / content_CVPR_2019 / html / Deng_ArcFace_Additive_Angular_Margin_Loss_for_Deep_Face_Recognition_CVPR_2019_paper.html>

[0007] Qiang Meng, et al., "MagFace: A Universal Representation for Face Recognition and Quality Assessment", [online], 2021 IEEE Conference Publication, [searched on December 20, 2022], Internet <URL: https: / / ieeexplore.ieee.org / document / <9578764>

[0008] Yifeng Chen, et al., "T-Center: A Novel Feature Extraction Approach Towards Large-Scale Iris Recognition", [online], February 12, 2020, IEEE, [searched on December 20, 2022], Internet <URL: https: / / ieeexplore.ieee.org / document / <8995585>

[0009] This disclosure aims to improve the technologies described in the above prior art documents.

[0010] According to one aspect of the present invention, there is provided an information processing system comprising: an extraction means for extracting iris information relating to an iris by using an extraction model that learns using a training image including an iris as an input; a calculation means for calculating a loss between the iris information and ground truth information using the resolution of the iris region included in the training image; and an update means for updating parameter values ​​of the extraction model using the calculated loss.

[0011] According to one aspect of the present invention, there is provided a learning device comprising: an extraction means for extracting iris information relating to an iris by using an extraction model as an input of a learning image including the iris; a calculation means for calculating a loss between the iris information and ground truth information using the resolution of the iris region included in the learning image; and an update means for updating parameter values ​​of the extraction model using the calculated loss.

[0012] According to one aspect of the present invention, there is provided a matching device comprising: an extraction means for matching that receives a target image including the iris of a target as an input, and extracts iris information relating to the iris of the target using the extraction model trained using the learning device according to claim 11; and a matching means that matches the iris information of the target with pre-registered registration data.

[0013] According to one aspect of the present invention, there is provided an information processing method in which a first computer group consisting of one or more computers extracts iris information related to the iris by using an extraction model that learns using a training image including the iris as input, calculates a loss between the iris information and ground truth information using the resolution of the iris region included in the training image, and updates parameter values ​​of the extraction model using the calculated loss.

[0014] According to one aspect of the present invention, there is provided an information processing method in which a second computer group consisting of one or more computers receives a target image including an iris of a target as an input, extracts iris information relating to the iris of the target using the extraction model learned by the first computer group by executing the information processing method of claim 13, and compares the iris information of the target with pre-registered registration data.

[0015] According to one aspect of the present invention, there is provided a recording medium having recorded thereon a program for causing a first computer group consisting of one or more computers to extract iris information relating to the iris by using an extraction model that learns using a training image including the iris as input, calculate a loss between the iris information and ground truth information using the resolution of the iris region included in the training image, and update parameter values ​​of the extraction model using the calculated loss.

[0016] According to one aspect of the present invention, there is provided a recording medium having recorded thereon a program for causing a second computer group consisting of one or more computers to extract iris information relating to the iris of the target using the extraction model learned by inputting a target image including the iris of the target and causing the first computer group to execute a program recorded on a recording medium described in claim 15, and to compare the iris information of the target with pre-registered registration data.

[0017] 1 is a diagram illustrating an overview of an information processing system according to embodiment 1. FIG. 2 is a diagram illustrating an overview of a learning device according to embodiment 1. FIG. 3 is a diagram illustrating an overview of a matching device according to embodiment 1. FIG. 4 is a flowchart illustrating an overview of a first example of information processing according to embodiment 1. FIG. 5 is a flowchart illustrating an overview of a second example of information processing according to embodiment 1. FIG. 6 is a diagram illustrating an example configuration of an information processing system according to embodiment 1. FIG. 7 is a diagram illustrating an example functional configuration of a learning device according to embodiment 1. FIG. 8 is a diagram illustrating an example of a learning image. FIG. 9 is a diagram illustrating an example physical configuration of a learning device 101 according to embodiment 1. FIG. 10 is a flowchart illustrating an example of learning processing according to embodiment 1. FIG. 11 is a flowchart illustrating an example of parameter calculation processing according to embodiment 1.

[0018] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In all the drawings, like components are designated by like reference numerals, and the description thereof will be omitted as appropriate.

[0019] 1 is a diagram showing an overview of an information processing system 100 according to embodiment 1. The information processing system 100 includes an extraction unit 112, a calculation unit 113, and an update unit 114.

[0020] The extraction unit 112 extracts iris information related to the iris by using an extraction model that learns using a learning image including the iris as an input.

[0021] The calculation unit 113 calculates the loss between the iris information and the correct information using the resolution of the iris region included in the learning image.

[0022] The update unit 114 updates the parameter values ​​of the extraction model using the calculated loss.

[0023] According to this information processing system 100, even if the resolution of the iris region in a target image including the iris of a target is low, it is possible to obtain highly accurate information from the image of the iris region.

[0024] 2 is a diagram illustrating an overview of the learning device 101 according to embodiment 1. The learning device 101 includes an extraction unit 112, a calculation unit 113, and an update unit 114.

[0025] The extraction unit 112 extracts iris information related to the iris by using an extraction model that learns using a learning image including the iris as an input.

[0026] The calculation unit 113 calculates the loss between the iris information and the correct information using the resolution of the iris region included in the learning image.

[0027] The update unit 114 updates the parameter values ​​of the extraction model using the calculated loss.

[0028] According to this learning device 101, even if the resolution of the iris region in a target image including the iris of the target is low, it is possible to obtain highly accurate information from the iris region.

[0029] 3 is a diagram showing an overview of the matching device 103 according to embodiment 1. The matching device 103 includes a matching extraction unit 131 and a matching unit 132.

[0030] The matching extraction unit 131 receives a target image including the iris of the target as input, and extracts iris information relating to the iris of the target using an extraction model trained using the learning device 101 described above.

[0031] The matching unit 132 matches the iris information of the target with pre-registered registration data.

[0032] According to this matching device 103, even if the resolution of the iris region in the target image including the iris of the target is low, it is possible to obtain highly accurate information from the iris region.

[0033] FIG. 4 is a flowchart illustrating an outline of a first example of information processing according to the first embodiment.

[0034] The extraction unit 112 extracts iris information related to the iris by using an extraction model that learns using a learning image including the iris as an input (step S102).

[0035] The calculation unit 113 calculates the loss between the iris information and the correct information using the resolution of the iris region included in the learning image (step S103).

[0036] The update unit 114 updates the parameter values ​​of the extraction model using the calculated loss (step S104).

[0037] According to this information processing, even if the resolution of the iris region in the target image including the iris of the target is low, it is possible to obtain highly accurate information from the iris region.

[0038] FIG. 5 is a flowchart illustrating an outline of a second example of information processing according to the first embodiment.

[0039] The matching extraction unit 131 receives a target image including the target iris as input, and extracts iris information related to the target iris using an extraction model learned by the learning device 101 executing the information processing method described above (step S201).

[0040] The matching unit 132 matches the iris information of the target with pre-registered registration data (step S202).

[0041] According to this information processing, even if the resolution of the iris region in the target image including the iris of the target is low, it is possible to obtain highly accurate information from the iris region.

[0042] A detailed example of the information processing system 100 according to the first embodiment will be described below.

[0043] (Details) Generally, when a target image including an iris is obtained by taking a photograph using a camera, the resolution of the iris region included in the target image may be small. The resolution of the iris region is the size of the iris region included in the target image. In other words, even if the resolution (number of pixels) of the target image is constant, the resolution of the iris region may vary.

[0044] When training a learning model to obtain iris information (information about the iris) from such target images, a general loss function such as that described in prior art documents is often used. However, in a learning model trained using such a loss function in a general manner, the accuracy of the iris information obtained from the target image may vary depending on the resolution of the iris region. In particular, the accuracy of iris information obtained from a target image with a low resolution of the iris region is often lower than the accuracy of iris information obtained from a target image with a high resolution of the iris region.

[0045] In view of the above-mentioned problems, one example of the object of the present invention is to provide an information processing system, a learning device, a matching device, an information processing method, a recording medium, etc. that solve the problem of obtaining accurate information from an image of an iris region in a target image including the iris of a target, even if the resolution of the iris region is low.

[0046] (Configuration Example of Information Processing System 100 According to First Embodiment) Fig. 6 is a diagram showing a configuration example of the information processing system 100 according to the first embodiment. The information processing system 100 is a system for matching a target using a target image and for learning an extraction model. The extraction model is a machine learning model that receives a target image as input and outputs information used for matching, etc.

[0047] The information processing system 100 includes a learning device 101 , an imaging device 102 , and a matching device 103 .

[0048] The learning device 101, the imaging device 102, and the matching device 103 are connected to one another via a network NT that is configured using a wired or wireless connection or a combination of these, and transmit and receive information to and from one another via the network NT.

[0049] The target image is an image that includes the iris of a target. The target is, for example, a person. Note that the target is not limited to a person, but may also be an animal such as a dog or a snake.

[0050] (Example of Functional Configuration of the Learning Device 101 According to Embodiment 1) Fig. 7 is a diagram showing an example of the functional configuration of the learning device 101 according to embodiment 1. The learning device 101 is a device for learning an extraction model. Functionally, the learning device 101 includes, for example, a learning information acquisition unit 111, an extraction unit 112, a calculation unit 113, and an update unit 114.

[0051] (Example of Function of Learning Information Acquisition Unit 111) The learning information acquisition unit 111 acquires learning information. The learning information is information for learning the extraction model. The learning information may be prepared in advance, for example.

[0052] The learning information includes, for example, a learning image, the resolution of the iris region, and correct answer information.

[0053] The training image is a target image used for training the extraction model. That is, the training image includes the iris of the target (training target) used for training the extraction model. The size (number of pixels) of the entire training image is, for example, a predetermined constant value. FIG. 8 is a diagram showing an example of a training image. Note that, although FIG. 8 shows an example in which the shape of the training image is rectangular, the shape may be changed as appropriate.

[0054] The iris region is a region showing the iris in the target image (including the learning image), that is, a partial region showing the iris in the target image. In Fig. 8, the iris region is indicated by a dot.

[0055] As described above, the resolution of the iris region is the size of the iris region. The size of the iris region is, for example, the number of pixels indicating the iris diameter. The iris diameter is the number of pixels of at least one of the diameter D, radius R, etc. of the outer edge of the iris region (i.e., the outer edge of the iris in the target image (including the learning image)) (see FIG. 8 ).

[0056] The resolution of the iris region included in the learning information is the resolution of the iris region included in the learning image.

[0057] The correct answer information includes the correct answer of the output value from the extraction model, that is, the correct answer of the information obtained from the target image using the extraction model.

[0058] The correct answer information may include, for example, information corresponding to at least a portion of the iris information output by the extraction model, such as at least one of the label or class of the iris included in the training image, an image of the iris region, and the position of the characteristic feature.

[0059] The label is information for identifying the class to which the iris included in the learning image belongs. The label is expressed using, for example, a predetermined index such as a letter, symbol, or number.

[0060] The class may be, for example, which iris of which eye of a learning object it is, left or right, etc. The class may be represented by, for example, a vector quantity such as a one-hot vector.

[0061] The characteristic location is at least one or more characteristic locations in a target image including the iris. The characteristic location may include at least one characteristic location such as the center of the pupil, the outer corner of the eye, the inner corner of the eye, the topmost point of the upper eyelid, or the bottommost point of the lower eyelid. Furthermore, the characteristic location is not limited to a characteristic location, and may include graphic information such as the radius of a circle or the length and width of a rectangle. Such characteristic locations may be determined in advance or automatically using a learning model.

[0062] In this embodiment, an example will be described in which the correct answer information includes a class.

[0063] (Example of Function of Extraction Unit 112) Referring again to FIG. 7 , the extraction unit 112 is configured to include an extraction model. The extraction model is a machine learning model for extracting iris information using a target image as input. That is, the extraction unit 112 according to this embodiment extracts iris information using an extraction model that learns using a learning image as input. The extraction unit 112 may output the extracted iris information to the calculation unit 113.

[0064] The iris information is information relating to the iris. The iris information includes, for example, at least one of a class, a label, an iris feature, an image of the iris region, and the position of a feature point. The iris feature is a feature extracted from an iris included in a target image (including a learning image). The iris feature is, for example, a vector quantity.

[0065] In this embodiment, an example will be described in which iris information includes iris feature amounts and classes.

[0066] In more detail, for example, the extraction model includes a feature extraction model and a classification model. Each of these models is a machine learning model configured using, for example, a neural network, and outputs extracted information. Corresponding to the configuration of such an extraction model, the extraction unit 112 functionally includes a feature extraction unit 112a and a classification unit 112b, as shown in FIG. 7 .

[0067] The feature extraction unit 112a is configured to include a feature extraction model. The feature extraction model is a machine learning model for extracting iris features from a target image. The feature extraction unit 112a extracts iris features using the feature extraction model with a training image as input.

[0068] The classification unit 112b is configured to include a classification model. The classification model is a machine learning model for extracting classes corresponding to irises included in a target image using iris feature amounts. The classification unit 112b uses the classification model with the iris feature amounts extracted from the training image by the feature amount extraction unit 112a as input to extract classes corresponding to irises included in the training image. Note that the classification model may, for example, extract labels instead of or in addition to classes.

[0069] (Example of Function of Calculation Unit 113) The calculation unit 113 calculates the loss between the iris information and the correct information using the resolution of the iris region included in the learning image.

[0070] For example, ArcFace or CosFace may be used as a loss function for calculating the loss. Note that the loss function is not limited to ArcFace or CosFace, and may be, for example, L2 softmax loss, cosine loss, MagFace, T-center, SphereFace, AdaCos, or the like.

[0071] In detail, for example, the calculation unit 113 functionally includes a parameter calculation unit 113a and a loss calculation unit 113b as shown in FIG.

[0072] The parameter calculation unit 113a calculates parameter values ​​related to parameters for calculating loss.

[0073] The parameter value includes, for example, at least one of the following (1) to (3): In this embodiment, an example in which the parameter value includes the following (1) to (3) will be described.

[0074] (1) Value of margin parameter (2) Value of weight decay parameter (3) Size of iris feature

[0075] Here, the margin parameter (1) is a hyperparameter included in the loss function for calculating the loss. Including the margin parameter value in the parameter value calculated by the parameter calculation unit 113a is suitable when, for example, ArcFace, CosFace, or the like is used as the loss function.

[0076] (2) The weight decay parameter is a hyperparameter of weight decay included in the loss function for calculating the loss.

[0077] (3) The magnitude of the iris feature amount is, for example, the L2 norm of the feature amount vector representing the iris feature amount.

[0078] The Lp norm is a positive value among the pth roots of the sum of the absolute values ​​of each component included in the feature vector raised to the pth power. For example, the L1 norm is the sum of the absolute values ​​of each component included in the feature vector. For example, the L2 norm is a Euclidean norm.

[0079] The parameter calculation unit 113a has functions for calculating, for example, each of (1) to (3). Fig. 9 is a diagram showing an example of the functional configuration of the parameter calculation unit 113a. Functionally, the parameter calculation unit 113a includes, for example, a margin calculation unit 113a_1, a weight decay calculation unit 113a_2, and a norm calculation unit 113a_3.

[0080] The margin calculation unit 113a_1 calculates a margin parameter using the resolution of the iris region. The weight attenuation calculation unit 113a_2 calculates a weight attenuation value using the resolution of the iris region. The norm calculation unit 113a_3 finds the magnitude of the iris feature amount.

[0081] Referring again to Fig. 7, the loss calculation unit 113b calculates the loss between the iris information and the correct information using the parameter values ​​calculated by the parameter calculation unit 113a.

[0082] (Function of Updater 114) The updater 114 uses the loss calculated by the calculator 113 to update the parameter values ​​of the extraction model.

[0083] In detail, as shown in FIG. 7, for example, the updating unit 114 functionally includes a gradient calculation unit 114a and a parameter updating unit 114b.

[0084] The gradient calculation unit 114a calculates the gradient for the parameters of the extraction model using the loss calculated by the loss calculation unit 113b.

[0085] The parameter update unit 114b updates the parameter values ​​of the extraction model using the gradient calculated by the gradient calculation unit 114a.

[0086] (Example of Functional Configuration of Image Capturing Device 102 According to First Embodiment) The image capturing device 102 is a camera or the like that captures an image of a target. For example, the image capturing device 102 generates an image of the target by capturing an image when receiving an image capturing instruction.

[0087] The shooting instructions may be output to the shooting device 102 from, for example, a sensor (not shown) for detecting that the subject is located in a predetermined shooting area, or a detection device (not shown) for detecting a face from an image of the subject photographed by another shooting device (not shown).

[0088] The target image only needs to include the iris of one of the target's eyes, and may be a binocular image including both of the target's eyes, a monocular image including either the target's left or right eye, a facial image including the target's face, or a whole-body image including the target's entire body.

[0089] (Example of Functional Configuration of the Verification Device 103 According to the First Embodiment) As described above with reference to FIG. 3, the verification device 103 includes the verification-use extraction unit 131 and the verification unit 132.

[0090] The extraction unit for matching 131 is configured to include, for example, an extraction model trained using the learning device 101. The extraction unit for matching 131 receives a target image including the iris of the target as input and extracts iris information of the target from the target image using the trained extraction model. The extraction unit for matching 131 may output the extracted iris information to the matching unit 132.

[0091] The matching unit 132 matches the iris information extracted by the matching extraction unit 131 with pre-registered registration data. The result of this matching can be used, for example, to authenticate whether or not the target has been pre-registered. However, the use of the result of matching is not limited to this.

[0092] Up to this point, the functional configuration example of the information processing system 100 according to the first embodiment has been mainly described. From here, a physical configuration example of the information processing system 100 according to this embodiment will be described.

[0093] (Physical configuration of information processing system 100) The information processing system 100 is physically composed of a learning device 101, an imaging device 102, and a matching device 103, which are connected via a network NT. Each of the learning device 101, the imaging device 102, and the matching device 103 is composed of, for example, a physically separate single device.

[0094] Note that some or all of the learning device 101, the imaging device 102, and the matching device 103 may be physically configured as a single device. Alternatively, the learning device 101, the imaging device 102, and the matching device 103 may be configured as multiple different devices connected via an appropriate communication line such as a network NT, for example, for each of one or more functions provided by each of them.

[0095] The first computer group may be composed of one or more devices having the functions of the learning device 101, and in this embodiment, is composed of the learning device 101. The second computer group may be composed of one or more devices having the functions of the matching device 103, and in this embodiment, is composed of the matching device 103.

[0096] The learning device 101 and the matching device 103 may have the same physical configuration. Here, an example of the physical configuration of the learning device 101 will be described with reference to the drawings.

[0097] 10 is a diagram showing an example of the physical configuration of the learning device 101 according to embodiment 1. The learning device 101 physically includes, for example, 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.

[0098] The bus 1010 is a data transmission path for transmitting and receiving data among the processor 1020, memory 1030, storage device 1040, user interface 1050, network interface 1060, camera 1070, and microphone 1080. However, the method of connecting the processor 1020 and the like to each other is not limited to bus connection.

[0099] The processor 1020 is implemented as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit).

[0100] The memory 1030 is a main storage device realized by a RAM (Random Access Memory) or the like.

[0101] 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 device that includes it (the learning device 101 in the example of FIG. 10). The processor 1020 loads each of these program modules into the memory 1030 and executes them to realize the function corresponding to that program module.

[0102] The network interface 1050 is an interface for connecting a device that includes it (the learning device 101 in the example of FIG. 10) to the network NT.

[0103] The input interface 1060 is an interface for the user to input information, and is configured from, for example, a touch panel, a keyboard, a mouse, and the like.

[0104] 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.

[0105] The photographing device 102 preferably includes a camera as described above. This camera preferably includes a network interface 1050 for connecting to the network NT.

[0106] So far, an example of the configuration of the information processing system 100 according to the first embodiment has been described. Next, the operation of the information processing system 100 according to this embodiment will be described.

[0107] (Operation of the Information Processing System 100 According to the First Embodiment) The information processing system 100 according to the first embodiment executes information processing including a learning process executed by the learning device 101 and a matching process executed by the matching device 103. Each of these processes will be described with reference to the drawings.

[0108] 11 is a flowchart illustrating an example of the learning process according to the first embodiment. The learning process is a process for learning an extraction model. The learning device 101 starts the learning process, for example, when the learning information acquisition unit 111 acquires learning information.

[0109] The learning information acquisition unit 111 acquires learning information of a mini-batch (step S101).

[0110] In detail, for example, the learning information acquisition unit 111 randomly acquires mini-batch learning information from the entire learning information acquired as an instruction to start the learning process. The mini-batch learning information is a part of the learning information and includes, for example, a number of learning images (e.g., 128 images) corresponding to a predetermined batch size, and resolution of the iris region and correct answer information corresponding to each learning image. The mini-batch learning information may include learning images with different resolutions of the iris region.

[0111] The correct answer information according to this embodiment includes, for example, the class of the iris contained in the learning image, as described above.

[0112] The batch size is not limited to 128 and may be determined in advance as appropriate. The learning information acquisition unit 111 may also perform an extension process on the learning images.

[0113] The extraction unit 112 extracts iris information using the extraction model with the learning image acquired in step S101 as an input (step S102).

[0114] In detail, for example, the feature extraction unit 112a extracts iris features using a feature extraction model with a learning image as input (step S102a).

[0115] The classification unit 112b uses the iris feature amount extracted in step S102a as an input and a classification model to extract a class corresponding to the iris included in the learning image (step S102b).

[0116] The calculation unit 113 calculates the loss between the iris information extracted in step S102 and the correct information obtained in step S101 using the resolution of the iris region obtained in step S101 (step S103).

[0117] In detail, for example, the parameter calculation unit 113a calculates parameter values ​​related to parameters for calculating loss (step S103a).

[0118] FIG. 12 is a flowchart showing an example of the parameter calculation process (step S103a) according to the first embodiment.

[0119] The margin calculation unit 113a_1 calculates the value of a margin parameter using the resolution of the iris region acquired in step S101 (step S103a_1).

[0120] For example, the lower the resolution of the iris region, the larger the value of the margin parameter that the margin calculation unit 113a_1 calculates. That is, the margin calculation unit 113a_1 increases the value of the margin parameter for a training image with a low resolution of the iris region, and decreases the value of the margin parameter for a training image with a high resolution of the iris region.

[0121] The weight attenuation calculation unit 113a_2 calculates the value of the weight attenuation parameter using the resolution of the iris region acquired in step S101 (step S103a_2).

[0122] For example, the lower the resolution of the iris region, the smaller the value of the weight attenuation parameter that the weight attenuation calculation unit 113a_2 calculates. That is, the weight attenuation calculation unit 113a_2 reduces the value of the weight attenuation parameter for training images with low-resolution iris regions and increases the value of the weight attenuation parameter for training images with high-resolution iris regions.

[0123] For example, the weight attenuation calculation unit 113a_2 calculates the value of the weight attenuation parameter wd(r) using wd(r) = wd0 * (min(r, 100) / 100). wd0 is the value of the weight attenuation parameter before the change, and is, for example, a constant. r is the resolution of the iris region (for example, the number of pixels of the iris diameter). min(a, b) is a function representing the smaller of a and b. In this formula, the value of the weight attenuation parameter wd(r) decreases in accordance with the resolution of the iris region when the resolution of the iris region is less than 100, and becomes a constant when the resolution of the iris region is 100 or greater.

[0124] The norm calculation unit 113a_3 calculates the size of the iris feature amount (step S103a_3). The size of the iris feature amount is, for example, the iris diameter (for example, the diameter or radius) of the iris included in the learning image.

[0125] Referring again to Fig. 11, the loss calculation unit 113b uses the parameter values ​​calculated in step S103a to calculate the loss between the iris information and the correct information (step S103b).

[0126] For example, a typical CosFace uses the losses expressed by equations (1) and (2), and a typical ArcFace uses the losses expressed by equation (3).

[0127]

[0128]

[0129]

[0130] where N is the batch size. ｙｉ，ｉ is the feature (F ｉ ) and the weight y ｉ-th column. S is a scaling parameter. m is a margin. W is a weight vector. The weight vector W is composed of components of the dimension of the number of classes CN x feature vector F, where F is a feature vector (vector quantity representing iris features). y is a one-hot vector representing the class (correct class) included in the correct answer information. By minimizing these losses, the distance between the weight of the positive example and the class (vector quantity) can be reduced.

[0131] In step S103b, such losses are corrected using the parameter values ​​determined in step S103a.

[0132] In detail, for example, the loss calculation unit 113b uses the value of the margin parameter calculated in step S103a_1 to calculate the loss.

[0133] For example, the loss calculation unit 113b uses the value of the weight decay parameter calculated in step S103a_2 to calculate the loss.

[0134] For example, when the resolution of the iris region acquired in step S101 is equal to or lower than a predetermined value, the loss calculation unit 113b calculates the loss so that the size of the iris feature amount calculated in step S103a_3 approaches a predetermined reference value. Here, the predetermined reference value is the size of the iris feature amount extracted from a training image having an iris region resolution of a predetermined value.

[0135] Here, if the reference value is D0 and the size of the iris feature calculated in step S103a_3 is D, for example, the loss calculation unit 113b further uses |max(D0-D), 0| as the loss E. max(A, B) is a function that represents the larger value of A and B.

[0136] The update unit 114 updates the parameter values ​​of the extraction model using the loss calculated in step S103 (step S104).

[0137] In detail, for example, the gradient calculation unit 114a calculates the gradient for the parameters of the extraction model using the loss calculated in step S103b (step S104a).

[0138] In step S104a, for example, the gradient calculation unit 114a calculates the gradient for the parameters of the extraction model by backpropagation using the loss calculated in step S103b.

[0139] The parameter update unit 114b updates the parameter values ​​of the extraction model using the gradient calculated in step S104a (step S104b), and ends the learning process.

[0140] In step S104b, for example, the parameter update unit 114b updates the parameters of the extraction model using the gradient calculated in step S104a and a preset learning rate.

[0141] This learning process may be repeated a predetermined number of times. The classes extracted in step S102b are random values ​​(e.g., vector quantities) in the early stages of learning, but as learning progresses, they become closer to the one-hot vector representing the correct class. This allows the extraction model to be learned.

[0142] In addition, if the change in loss calculated in step S103 becomes equal to or less than a predetermined threshold, the loss calculation unit 113b may terminate the learning process even if the number of times the learning process has been repeated is less than the predetermined number.

[0143] (Example of Matching Process According to Embodiment 1) The matching process according to embodiment 1 includes, for example, the process described above with reference to Fig. 5. The matching process is a process for matching a target using a target image and a trained extraction model (or a feature extraction model). The matching device 103 starts the matching process, for example, when the matching extraction unit 131 acquires a target image including the iris of the target.

[0144] The matching extraction unit 131 receives a target image including the iris of the target as input and uses an extraction model to extract iris information of the target from the target image (step S201).

[0145] The extraction model used here is an extraction model that has been learned by repeatedly executing the learning process. Note that, in step S201, only the feature extraction model may be applied.

[0146] The iris information extracted in step S201 may include, for example, iris feature amounts.

[0147] The matching unit 132 matches the iris information extracted in step S201 with pre-registered data (step S202), and ends the matching process.

[0148] For example, the matching unit 132 may output, as a result of the matching, the similarity between the extracted iris feature amount and the iris feature amount included in the enrollment data, etc. Furthermore, for example, the matching unit 132 may output, as a result of the matching, information for identifying an individual, such as an individual ID (Identifier) ​​assigned to each individual and an individual name, which are included in the enrollment data, for enrollment data in which the similarity between the iris feature amount included in the enrollment data is equal to or greater than a threshold value or is the largest.

[0149] (Operations and Effects) As described above, according to this embodiment, the learning device 101 includes the extraction unit 112, the calculation unit 113, and the update unit 114. The extraction unit 112 extracts iris information related to the iris by using an extraction model that learns using training images including irises as input. The calculation unit 113 calculates the loss between the iris information and the correct information using the resolution of the iris region included in the training images. The update unit 114 updates the parameter values ​​of the extraction model using the calculated loss.

[0150] This allows the parameter values ​​of the extraction model to be updated using a loss that corresponds to the resolution of the iris region included in the training image, making it possible to obtain highly accurate information from the image of the iris region even if the resolution of the iris region in the target image containing the target iris is low.

[0151] According to this embodiment, the calculation unit 113 includes a parameter calculation unit 113 a and a loss calculation unit 113 b. The parameter calculation unit 113 a calculates parameter values ​​related to parameters for calculating losses. The loss calculation unit 113 b uses the parameter values ​​to calculate losses between the iris information and the correct information.

[0152] This allows the parameter values ​​of the extraction model to be updated using a loss that corresponds to the resolution of the iris region included in the training image, making it possible to obtain highly accurate information from the image of the iris region even if the resolution of the iris region included in the target image is low.

[0153] According to this embodiment, the parameter values ​​include at least one of (1) the value of a margin parameter, which is a hyperparameter included in a loss function for calculating a loss, (2) the value of a weight decay parameter, and (3) the magnitude of the iris feature.

[0154] By including the value of the margin parameter (1) in the parameter values, it is possible to construct an extraction model that has excellent discrimination performance for target images containing the irises of different people, even if the resolution of the iris region contained in the target image is low.

[0155] Similarly, by including the parameter value (2) the value of the weight decay parameter, an extraction model can be constructed that has excellent discrimination performance for target images containing irises of different people, even if the resolution of the iris region contained in the target image is low.

[0156] By including the parameter value (3) the magnitude of the iris feature amount, even if the resolution of the iris region included in the target image is low, it is possible to obtain a loss according to the resolution of the iris region.

[0157] Therefore, even if the resolution of the iris region in a target image including the iris of the target is low, it is possible to obtain highly accurate information from the image of the iris region.

[0158] According to this embodiment, the parameter values ​​include (1) the value of a margin parameter, which is a hyperparameter included in the loss function. The parameter calculation unit 113a includes a margin calculation unit 113a_1 that calculates a larger value for the margin parameter as the resolution of the iris region decreases.

[0159] By setting constraints on the margin parameters in this way, it is possible to construct an extraction model that has excellent discrimination performance for target images containing the irises of different people, even if the resolution of the iris region included in the target image is low. Therefore, even if the resolution of the iris region in the target image containing the target iris is low, it is possible to obtain accurate information from the image of the iris region.

[0160] According to this embodiment, the parameter values ​​include (2) the value of a weight decay parameter. The parameter calculation unit 113a includes a weight decay calculation unit 113a_2 that determines a smaller value for the weight decay parameter as the resolution of the iris region decreases.

[0161] By restricting the value of the weight decay parameter in this way, it is possible to construct an extraction model that has excellent discrimination performance for target images containing irises of different people, even if the resolution of the iris region included in the target image is low. Therefore, even if the resolution of the iris region in the target image containing the target iris is low, it is possible to obtain accurate information from the image of the iris region.

[0162] According to this embodiment, the parameter values ​​include (3) the magnitude of the iris feature amount. The parameter calculation unit 113a includes a norm calculation unit 113a_3 that calculates the magnitude of the iris feature amount. The loss calculation unit 113b calculates the loss between the iris information and the correct information so that the magnitude of the iris feature amount approaches a predetermined reference value when the resolution of the iris region is equal to or lower than a predetermined value. The predetermined reference value is the magnitude of the iris feature amount extracted from a training image having an iris region with a predetermined resolution.

[0163] This allows for constraints to be set so that the resolution of the iris region in the training image remains constant when the resolution of the iris region is low. Therefore, even when the resolution of the iris region in the target image is low, an extraction model with excellent discrimination performance for target images containing the irises of different people can be constructed. Therefore, even when the resolution of the iris region in the target image containing the iris of the target is low, it is possible to obtain highly accurate information from the image of the iris region.

[0164] According to this embodiment, the iris information includes at least one of iris features extracted from the training image, an image of the iris region in the training image, the position of the characteristic location, and a class corresponding to the iris included in the training image.

[0165] This allows various pieces of iris information to be obtained using the extraction model, making it possible to obtain various pieces of highly accurate information from the image of the iris region even if the resolution of the iris region included in the target image is low.

[0166] (Variation 1) As described above, the parameter values ​​may include at least one of (1) the value of the margin parameter, (2) the value of the weight decay parameter, and (3) the size of the iris feature amount. In other words, the parameter values ​​may be any one or two of (1) to (3).

[0167] In this case, the parameter calculation unit 113a may include one or more of a margin calculation unit 113a_1, a weight attenuation calculation unit 113a_2, and a norm calculation unit 113a_3 for calculating the parameter value (1) to (3) depending on whether the parameter value includes the parameter value (1) to (3). Furthermore, the parameter calculation process (step S103a) may include one or more of steps 103a_1 to 103a_3 for calculating the parameter value (1) to (3) depending on whether the parameter value includes the parameter value (1). In step S103b, the loss may be corrected using the parameter value calculated in any of steps 103a_1 to 103a_3.

[0168] This modification also provides the same effects as the first embodiment.

[0169] (Modification 2) In this modification, an example of a loss function for calculating a loss using the resolution of the iris region included in the learning image when the iris information includes an image of the iris region will be described.

[0170] The loss E2 in this modified example is, for example, the loss E (see Equation 4) obtained by using the L1 norm between an extraction map showing an image of the iris region and a correct answer map showing an image of the iris region included in the correct answer information, divided by the resolution of the iris region included in the training image (for example, the number of pixels in the iris diameter).

[0171]

[0172] Here, j is information (image ID) for identifying the learning image included in the batch, and M is the number of samples. ｊ is the resolution of the iris region included in the j-th learning image. i is a value indicating each pixel. N is the total number of pixels in the learning image. y ｉ is the pixel value of the correct map. x is a vector quantity composed of pixel values ​​of the learning image. f(x) ｉ is the pixel value of the extraction map. Note that the loss E2 is not limited to this, and may be, for example, other than the L1 norm.

[0173] This modification also provides the same effects as the first embodiment.

[0174] (Modification 3) In this modification, an example of a loss function for calculating a loss using the resolution of the iris region included in the learning image when the iris information includes the positions of characteristic parts will be described.

[0175] The loss E3 in this modification is, for example, a value obtained by dividing the loss E obtained by using the L1 norm between the extracted position, which is the position of the extracted characteristic feature, and the correct position, which indicates the position of the characteristic feature included in the correct answer information, by the resolution of the iris region included in the training image (for example, the number of pixels in the iris diameter). The loss E here may be the loss E calculated using equation (4) as in modification 2.

[0176] In this modification, j is information (image ID) for identifying the learning image included in the batch, and M is the number of samples. ｊ is the resolution of the iris region included in the j-th learning image. i is a value indicating each pixel. N is the total number of pixels in the learning image. y ｉ is the correct position. x is a vector quantity composed of pixel values ​​of the learning image. f(x) ｉ is the extracted position. For example, if there are three feature locations, the elements that make up those locations are three x-coordinates and three y-coordinates, for a total of six elements. Note that the loss E3 is not limited to this, and may be, for example, other than the L1 norm.

[0177] This modification also provides the same effects as the first embodiment.

[0178] (Embodiment 2) In embodiment 2, an example will be described in which the parameter values ​​are changed when the parameter values ​​include (3) the size of the iris feature amount. In this embodiment, for simplicity, descriptions that overlap with embodiment 1 will be omitted as appropriate.

[0179] The parameter values ​​according to this embodiment include (3) the magnitude of an iris feature amount. The parameter values ​​according to this embodiment differ from those of the first embodiment in that they include the magnitudes of a plurality of iris feature amounts. The plurality of iris feature amounts include, for example, low-resolution feature amounts and high-resolution feature amounts extracted from low-resolution training images and high-resolution training images that share common ground information but have different resolutions for the iris region.

[0180] The low-resolution training image and the high-resolution training image are training images with low and high resolutions of the iris region, respectively. The low-resolution feature and the high-resolution feature are iris feature extracted from the low-resolution training image and the high-resolution training image, respectively.

[0181] The norm calculation unit 113a_3 according to this embodiment may calculate the magnitude of each of the low-resolution feature amount and the high-resolution feature amount.

[0182] The loss calculation unit 113b according to this embodiment may calculate the loss between the iris information and the correct information so that the magnitude of the low-resolution feature approaches the magnitude of the high-resolution feature. Here, if the magnitude of the high-resolution feature is DH and the magnitude of the low-resolution feature is DL, for example, the loss calculation unit 113b may further use |DH-DL| as the loss E.

[0183] Note that using this loss may cause the magnitude DH of the high-resolution feature to approach the magnitude DL of the low-resolution feature. To reduce this possibility, a minimum value D0 may be set in advance for the magnitude DH of the high-resolution feature vector. In this case, for example, |max(DH, D0)-DL| may be further used as the loss E.

[0184] As described above, according to this embodiment, the parameter value includes the magnitudes of a plurality of iris feature amounts. The plurality of iris feature amounts include low-resolution feature amounts and high-resolution feature amounts extracted from low-resolution training images and high-resolution training images that share common ground information but have different resolutions for the iris region.

[0185] The parameter calculation unit 113a includes a norm calculation unit 113a_3 that calculates the magnitude of each of the low-resolution feature amount and the high-resolution feature amount. The loss calculation unit 113b calculates the loss between the iris information and the correct information so that the magnitude of the low-resolution feature amount approaches the magnitude of the high-resolution feature amount.

[0186] By imposing such constraints on the loss, it is possible to train an extraction model so that iris information extracted from a low-resolution image of the iris region approaches iris information extracted from a high-resolution image of the iris region. Therefore, even if the resolution of the iris region included in the target image is low, it is possible to construct an extraction model with excellent performance for discriminating between target images containing the irises of different people. Therefore, even if the resolution of the iris region in the target image containing the target iris is low, it is possible to obtain accurate information from the image of the iris region.

[0187] Third Embodiment In a third embodiment, an example of changing a learning rate for updating parameters of an extraction model will be described. In this embodiment, for simplicity of explanation, explanations that overlap with those in the first embodiment will be omitted as appropriate.

[0188] The parameter update unit 114b according to this embodiment applies different learning rates to the feature extraction model and the classification model. For example, the learning rate of the feature extraction model may be a real number multiple (where this real number is greater than 1) or an integer multiple such as 2, 3, 4, or 5 of the learning rate of the classification model.

[0189] In the parameter update process (step S104b) according to this embodiment, the parameter update unit 114b may update the parameters of the extraction model by using the gradient calculated in step S104a and different learning rates for the feature extraction model and the classification model.

[0190] (Operations and Effects) As described above, according to this embodiment, the extraction model includes a feature extraction model that extracts iris feature amounts using training images as input, and a classification model that extracts classes corresponding to irises included in the training images using the iris feature amounts as input. The update unit 114 applies different learning rates to the feature extraction model and the classification model.

[0191] This allows the feature extraction model and the classification model to be trained at different amplitudes. For example, the feature extraction model can be trained at a larger amplitude than the classification model. Because the magnitude of the iris feature increases, the effect of training using the loss calculated by the loss calculation unit 113b can be further increased, making it more difficult for the magnitude of the high-resolution feature to decrease and more likely for the magnitude of the low-resolution feature to increase.

[0192] Therefore, even if the resolution of the iris region included in the target image is low, it is possible to construct an extraction model with excellent discrimination performance for target images containing the irises of different people. Therefore, even if the resolution of the iris region in the target image containing the iris of the target is low, it is possible to obtain highly accurate information from the image of the iris region.

[0193] Fourth Embodiment In a fourth embodiment, an example in which the gradient of a feature extraction model is multiplied by a constant will be described. In this embodiment, for simplicity, descriptions that overlap with those in the first embodiment will be omitted as appropriate.

[0194] The gradient calculation unit 114a according to this embodiment multiplies the gradient calculated for the feature extraction model by a constant. This constant may be a real number greater than 1, or an integer such as 2, 3, 4, or 5.

[0195] In the gradient calculation process (step S104a) according to this embodiment, the gradient calculation unit 114a calculates the gradients of the feature extraction model and the classification model with respect to their respective parameters. Then, the gradient calculation unit 114a multiplies the gradients calculated for the feature extraction model by a constant.

[0196] In the parameter update process (step S104b) according to this embodiment, the parameter update unit 114b updates the parameters of the feature extraction model using the gradient multiplied by a constant. As with embodiment 1, the parameter update unit 114b may update the parameters of the extraction model using the gradient calculated in step S104a. Furthermore, in this embodiment, the learning rates of the feature extraction model and the classification model may be the same as in embodiment 1, or may be different as in embodiment 3.

[0197] (Actions and Effects) As described above, according to this embodiment, the extraction model includes a feature extraction model that extracts iris feature amounts using a training image as input, and a classification model that extracts a class corresponding to the iris included in the training image using the iris feature amounts as input. The update unit 114 includes a gradient calculation unit 114a and a parameter update unit 114b. The gradient calculation unit 114a calculates a gradient for the parameters of the extraction model using the calculated loss. The parameter update unit 114b calculates and updates the parameters of the extraction model using the calculated gradient. In calculating the parameters of the extraction model, the parameter update unit 114b uses a value obtained by multiplying the calculated gradient by a predetermined constant.

[0198] As a result, similar to the third embodiment, the feature extraction model and the classification model can be trained at different amplitudes. For example, the feature extraction model can be trained at a larger amplitude than the classification model. Because the magnitude of the iris feature increases, the effect of loss correction can be further increased, making it difficult for the magnitude of the high-resolution feature to decrease and easy for the magnitude of the low-resolution feature to increase.

[0199] Therefore, even if the resolution of the iris region included in the target image is low, it is possible to construct an extraction model with excellent discrimination performance for target images containing the irises of different people. Therefore, even if the resolution of the iris region in the target image containing the iris of the target is low, it is possible to obtain highly accurate information from the image of the iris region.

[0200] Although the embodiments and modifications of the present invention have been described above with reference to the drawings, these are merely examples of the present invention, and various configurations other than those described above can also be adopted.

[0201] In addition, although the flowcharts used in the above description show multiple steps (processes) in a sequential order, the order of steps performed in each embodiment is not limited to the order shown. In each embodiment, the order of steps shown in the drawings can be changed as long as it does not cause any problems in terms of content. Furthermore, the above-described embodiments and variations can be combined as long as the content is not contradictory.

[0202] Some or all of the above-described embodiments can be described as, but are not limited to, the following supplementary notes.

[0203] 1. An information processing system comprising: extraction means for extracting iris information related to an iris by using an extraction model that learns using a training image including the iris as input; calculation means for calculating a loss between the iris information and correct information using the resolution of an iris region included in the training image; and update means for updating parameter values ​​of the extraction model using the calculated loss. 2. The information processing system described in 1., wherein the calculation means includes: parameter calculation means for calculating a parameter value related to a parameter for calculating the loss; and loss calculation means for calculating the loss between the iris information and correct information using the parameter value. 3. The information processing system described in 2., wherein the parameter value includes at least one of (1) the value of a margin parameter that is a hyperparameter included in a loss function for calculating the loss, (2) the value of a weight decay parameter included in the loss function for calculating the loss, and (3) the magnitude of the iris feature amount. 4. The parameter value includes the value of a margin parameter that is a hyperparameter included in the loss function, and the parameter calculation means includes margin calculation means that calculates a larger value for the margin parameter the lower the resolution of the iris region. 3. 5. The information processing system described in 3. or 4., wherein the parameter values ​​include the value of the weight decay parameter, and the parameter calculation means includes weight decay calculation means that sets a smaller value for the weight decay parameter as the resolution of the iris region becomes lower. 6. The information processing system described in any one of 3. to 5., wherein the parameter values ​​include the magnitude of the iris feature amount, and the parameter calculation means includes norm calculation means that calculates the magnitude of the iris feature amount, and the loss calculation means calculates the loss between the iris information and correct information so that the magnitude of the iris feature amount approaches a predetermined reference value when the resolution of the iris region is equal to or lower than a predetermined value, and the predetermined reference value is the magnitude of the iris feature amount extracted from the training image in which the resolution of the iris region is a predetermined value.7. The information processing system described in any one of 3. to 5., wherein the parameter value includes magnitudes of the plurality of iris feature amounts, the plurality of iris feature amounts include low-resolution feature amounts and high-resolution feature amounts extracted from low-resolution training images and high-resolution training images having the same correct answer information but different resolutions of the iris region, respectively, the parameter calculation means includes norm calculation means for calculating the magnitudes of the low-resolution feature amounts and the high-resolution feature amounts, and the loss calculation means calculates the loss between the iris information and correct answer information so that the magnitude of the low-resolution feature amount approaches the magnitude of the high-resolution feature amount. 8. The information processing system described in any one of 1. to 7., wherein the iris information includes at least one of iris feature amounts extracted from the training images, an image of the iris region in the training images, positions of characteristic locations, and a class corresponding to the iris included in the training images. 10. The information processing system according to any one of 1. to 8., wherein the extraction model includes: a feature extraction model that takes the training image as input and extracts the iris feature, and a classification model that takes the iris feature as input and extracts a class corresponding to the iris included in the training image, and the updating means applies different learning rates to the feature extraction model and the classification model. 11. The information processing system according to any one of 1. to 8., wherein the extraction model includes: a feature extraction model that takes the training image as input and extracts the iris information, and a classification model that takes the iris information as input and extracts a class corresponding to the iris included in the training image, and the updating means includes: gradient calculation means that uses the calculated loss to calculate a gradient for parameters of the extraction model, and parameter update means that uses the calculated gradient to calculate and update parameters of the extraction model, and the parameter update means uses a value obtained by multiplying the calculated gradient by a predetermined constant in calculating the parameters of the extraction model.11. A learning device comprising: extraction means for extracting iris information about an iris by using an extraction model with a learning image including the iris as input; calculation means for calculating a loss between the iris information and correct information using the resolution of the iris region included in the learning image; and update means for updating parameter values ​​of the extraction model using the calculated loss. 12. A matching device comprising: extraction means for matching that extracts iris information about the iris of a target by using the extraction model trained with the learning device described in 11 with a target image including the iris as input; and matching means for matching the iris information of the target with pre-registered enrollment data. 13. An information processing method in which a first computer group consisting of one or more computers extracts iris information about the iris by using an extraction model trained with a learning image including the iris as input; calculates the loss between the iris information and correct information using the resolution of the iris region included in the learning image; and updates parameter values ​​of the extraction model using the calculated loss. 14. An information processing method in which a second computer group consisting of one or more computers extracts iris information related to the iris of a target using the extraction model trained by the first computer group executing the information processing method described in 13. as input, and compares the iris information of the target with pre-registered registration data. 15. A recording medium having recorded thereon a program for causing a first computer group consisting of one or more computers to extract iris information related to the iris using an extraction model trained using a training image including an iris as input, calculate a loss between the iris information and correct information using the resolution of the iris region included in the training image, and update parameter values ​​of the extraction model using the calculated loss.16. A recording medium having recorded thereon a program for causing a second computer group consisting of one or more computers to extract iris information relating to the iris of the target using the extraction model learned by having the first computer group execute the program recorded on the recording medium described in 15. as input of a target image including the iris of the target, and compare the iris information of the target with pre-registered registration data.

[0204] REFERENCE SIGNS LIST 100 Information processing system 101 Learning device 102 Imaging device 103 Matching device 111 Learning information acquisition unit 112 Extraction unit 112a Feature extraction unit 112b Classification unit 113 Calculation unit 113a Parameter calculation unit 113a_1 Margin calculation unit 113a_2 Weight decay calculation unit 113a_3 Norm calculation unit 113b Loss calculation unit 114 Update unit 114a Gradient calculation unit 114b Parameter update unit 131 Matching extraction unit 132 Matching unit

Claims

1. an extraction means for extracting iris information relating to the iris by using an extraction model that learns using a learning image including the iris as an input; a calculation means for calculating a loss between the iris information and correct information using a resolution of an iris region included in the learning image; and an update means for updating the parameter values ​​of the extraction model using the calculated loss. Information processing system.

2. The calculation means a parameter calculation means for calculating a parameter value related to a parameter for calculating the loss; a loss calculation means for calculating a loss between the iris information and correct information using the parameter value; The information processing system according to claim 1 .

3. The parameter values ​​include at least one of: (1) a value of a margin parameter, which is a hyperparameter included in a loss function for calculating the loss; (2) a value of a weight decay parameter included in a loss function for calculating the loss; and (3) a size of an iris feature amount. The information processing system according to claim 2 .

4. the parameter values ​​include a value of a margin parameter, which is a hyperparameter included in the loss function; The parameter calculation means includes margin calculation means for calculating a larger value for the margin parameter as the resolution of the iris region decreases. The information processing system according to claim 3 .

5. the parameter values ​​include the value of the weight decay parameter; The parameter calculation means includes weight decay calculation means that determines a smaller value for the weight decay parameter as the resolution of the iris region decreases.

5. The information processing system according to claim 3 or 4.

6. the parameter value includes the magnitude of the iris feature amount, the parameter calculation means includes a norm calculation means for calculating a magnitude of the iris feature amount, the loss calculation means calculates a loss between the iris information and the correct information so that a size of the iris feature amount approaches a predetermined reference value when a resolution of the iris area is equal to or less than a predetermined value; The predetermined reference value is the size of the iris feature amount extracted from the learning image in which the resolution of the iris region is a predetermined value.

5. The information processing system according to claim 3 or 4.

7. the parameter values ​​include the magnitudes of the plurality of iris feature amounts, the plurality of iris features include low-resolution features and high-resolution features extracted from low-resolution training images and high-resolution training images having the same correct answer information but different resolutions of the iris region, the parameter calculation means includes norm calculation means for calculating the magnitude of each of the low-resolution feature amount and the high-resolution feature amount; The loss calculation means calculates a loss between the iris information and the correct information so that the magnitude of the low-resolution feature amount approaches the magnitude of the high-resolution feature amount.

5. The information processing system according to claim 3 or 4.

8. The iris information includes at least one of an iris feature amount extracted from the training image, an image of an iris region in the training image, a position of a feature point, and a class corresponding to the iris included in the training image. The information processing system according to any one of claims 1 to 4.

9. a first computer group consisting of one or more computers, extracting iris information about the iris by using an extraction model that learns using a learning image including the iris as an input; Calculating a loss between the iris information and the correct information using a resolution of the iris region included in the learning image; The calculated loss is used to update the parameter values ​​of the extraction model. Information processing methods.

10. a first computer group consisting of one or more computers; extracting iris information about the iris by using an extraction model that learns using a learning image including the iris as an input; Calculating a loss between the iris information and the correct information using a resolution of the iris region included in the learning image; a program for executing updating of parameter values ​​of the extraction model using the calculated loss;

Images