Information processing systems, information processing methods, and programs

Abstract

An information processing system (100) comprises a quality estimation unit (111) and an image processing unit (113). The quality estimation unit (111) estimates quality information of an image of interest. The image processing unit (113) processes the image of interest by using a neural network corresponding to a network policy which is determined on the basis of quality information of the image of interest.

Description

Technical Field

[0001] This disclosure relates to an information processing system , love an information processing method, and program is related thereto.

Background Art

[0002] In iris authentication, generally, when the quality such as the resolution of an image is low, the authentication accuracy may decrease. Therefore, in iris authentication, it is desirable to capture an authentication target using a high-resolution imaging device, but it may be difficult to use a high-resolution imaging device due to installation space, cost, etc. of the imaging device.

[0003] For example, Patent Document 1 discloses that when performing iris authentication, a predetermined number of captured images are used to generate an iris authentication image with a second resolution higher than the first resolution of the captured image. As an example thereof, Patent Document 1 discloses that from eight continuously captured images, 1×1 pixels of the captured image are converted into 2×2 pixels to generate an authentication image with a resolution twice as high.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] This disclosure aims to improve the technology described in the above-mentioned prior art documents.

Means for Solving the Problems

[0006] According to one aspect of the present invention, quality estimation means for estimating quality information of a target image, Image processing means for processing the target image using a neural network according to a network policy determined based on the quality information of the target image 、 A learning means that trains the neural network by inputting training target images into the neural network, A learning policy determination means for training a first policy determination model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, A first loss acquisition means for determining the first loss when processing the training target image using a neural network corresponding to each of the multiple network policies, and is provided 、 The aforementioned first policy decision model consists of a neural network, The aforementioned learning policy determination means is: A first learning likelihood acquisition means inputs the quality information of the target images for training into the first policy decision model and calculates the likelihood of each network policy, The system includes a second parameter updating means for updating the parameters of the first policy decision model based on the first loss and likelihood of each of the aforementioned network policies, The second parameter updating means is A second loss acquisition means for obtaining a second loss which is the sum of the first losses weighted by likelihood for each of the aforementioned network policies, Includes a second parameter update means for updating the parameters of the first policy decision model based on the second loss. An information processing system is provided. According to one aspect of the present invention, A quality estimation means for estimating the quality information of the target image, Image processing means that processes the target image using a neural network according to a network policy determined based on the quality information of the target image, A learning means that trains the neural network by inputting training target images into the neural network, The system includes a learning policy determination means for training a third policy determination model for determining the network policy corresponding to the target image from among a predetermined number of network policies, The aforementioned learning policy determination means is: A means for obtaining a correct likelihood by inputting the quality information of the training target image into a trained first policy decision model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, and obtaining the likelihood of each network policy as the correct likelihood, A third learning likelihood acquisition means inputs the aforementioned training target images into a third policy decision model and obtains the likelihood of each network policy as the likelihood during training, The system includes a fourth parameter update means for updating the parameters of the third policy decision model based on the fourth loss such that the likelihood during learning approaches the true likelihood, The fourth loss is a loss based on the error between the likelihood during learning and the true likelihood. An information processing system is provided.

[0008] According to one aspect of the present invention one or more computers estimate the quality information of the target image process the target image using a neural network according to a network policy determined based on the quality information of the target image death, By inputting the training target images into the neural network, the neural network is trained. A first policy decision model is trained to determine the network policy corresponding to the quality information of the target image from among a predetermined number of network policies. The first loss is calculated when processing the training target image using a neural network corresponding to each of the aforementioned multiple network policies. including fruit, The aforementioned first policy decision model consists of a neural network, Training the aforementioned first policy decision model involves, The quality information of the training target images is input into the first policy decision model to obtain the likelihood of each network policy. This includes updating the parameters of the first policy decision model based on the first loss and likelihood of each of the aforementioned network policies, Updating the parameters of the first policy decision model mentioned above means For each of the aforementioned network policies, calculate the second loss, which is the sum of the first losses weighted by likelihood. This includes updating the parameters of the first policy decision model based on the second loss. Information processing methods are provided. According to one aspect of the present invention, One or more computers, Estimate the quality information of the target image, The target image is processed using a neural network that corresponds to a network policy determined based on the quality information of the target image. By inputting the training target images into the neural network, the neural network is trained. This includes training a third policy decision model for determining the network policy corresponding to the target image from among a predetermined number of network policies, Training the aforementioned third policy decision model is The quality information of the target image for training is input into a trained first policy decision model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, and the likelihood of each network policy is obtained as the correct likelihood. The aforementioned training target images are input into the third policy decision model, and the likelihood of each network policy is obtained as the likelihood during training. This includes updating the parameters of the third policy decision model based on the fourth loss so that the likelihood during learning approaches the true likelihood, The fourth loss is a loss based on the error between the likelihood during learning and the true likelihood. Information processing methods are provided.

[0009] According to one aspect of the present invention, On one or more computers, Estimate the quality information of the target image, The target image is processed using a neural network that corresponds to a network policy determined based on the quality information of the target image. death, By inputting the training target images into the neural network, the neural network is trained. A first policy decision model is trained to determine the network policy corresponding to the quality information of the target image from among a predetermined number of network policies. The first loss is calculated when processing the training target image using a neural network corresponding to each of the aforementioned multiple network policies. Make them do it 、 The aforementioned first policy decision model consists of a neural network, Training the aforementioned first policy decision model involves, The quality information of the training target images is input into the first policy decision model to obtain the likelihood of each network policy. This includes updating the parameters of the first policy decision model based on the first loss and likelihood of each of the aforementioned network policies, Updating the parameters of the first policy decision model mentioned above means For each of the aforementioned network policies, calculate the second loss, which is the sum of the first losses weighted by likelihood. This includes updating the parameters of the first policy decision model based on the second loss, Professional Mu It will be provided. According to one aspect of the present invention, On one or more computers, Estimate the quality information of the target image, The target image is processed using a neural network that corresponds to a network policy determined based on the quality information of the target image. By inputting the training target images into the neural network, the neural network is trained. The system is made to perform training on a third policy decision model for determining the network policy corresponding to the target image from among a predetermined number of network policies. Training the aforementioned third policy decision model is The quality information of the target image for training is input into a trained first policy decision model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, and the likelihood of each network policy is obtained as the correct likelihood. The aforementioned training target images are input into the third policy decision model, and the likelihood of each network policy is obtained as the likelihood during training. This includes updating the parameters of the third policy decision model based on the fourth loss so that the likelihood during learning approaches the true likelihood, A program is provided in which the fourth loss is a loss based on the error between the likelihood during learning and the true likelihood. [Brief explanation of the drawing]

[0010] [Figure 1] This diagram shows an overview of the information processing system according to Embodiment 1. [Figure 2] This diagram shows an overview of the information processing device according to Embodiment 1. [Figure 3] This is a flowchart outlining the information processing according to Embodiment 1. [Figure 4] This figure shows an example configuration of the information processing system according to Embodiment 1. [Figure 5] This figure shows an example of the neural network configuration according to Embodiment 1. [Figure 6] This figure shows an example of the physical configuration of the information processing device according to Embodiment 1. [Figure 7] This flowchart shows an example of information processing according to Embodiment 1. [Figure 8] This figure shows an example of a target image related to Embodiment 1. [Figure 9] This figure shows an example of the functional configuration of the image processing unit according to Embodiment 2. [Figure 10] This figure shows an example of the neural network configuration according to Embodiment 2. [Figure 11] This flowchart shows a detailed example of image processing according to Embodiment 2. [Figure 12] This figure shows an example of the iris region according to Embodiment 2. [Figure 13] This figure shows an example of a standard image relating to Embodiment 2. [Figure 14] This figure shows an example of the configuration of a neural network according to Embodiment 3. [Figure 15] This figure shows an example configuration of the information processing system according to Embodiment 4. [Figure 16] This flowchart shows an example of the learning process according to Embodiment 4. [Figure 17] This figure shows an example of a learning probability distribution according to Embodiment 4. [Figure 18] This figure shows an example configuration of the information processing system according to Embodiment 5. [Figure 19] This flowchart shows an example of the learning process according to Embodiment 5. [Figure 20] This figure shows an example of a selection probability distribution according to Embodiment 5. [Figure 21] This figure shows an example configuration of the information processing system according to Embodiment 6. [Figure 22] This flowchart shows an example of the learning process according to Embodiment 6. [Figure 23] This figure shows an example configuration of the information processing system according to Embodiment 7. [Figure 24] This is a flowchart showing an example of information processing according to Embodiment 7. [Figure 25] This flowchart shows an example of the network policy determination process according to Embodiment 7. [Figure 26] This figure shows an example configuration of the information processing system according to Embodiment 8. [Figure 27] This flowchart shows an example of the learning process according to Embodiment 8. [Figure 28] This figure shows an example configuration of the learning policy determination unit according to Embodiment 9. [Figure 29] This flowchart shows an example of the first policy decision model learning process according to Embodiment 9. [Figure 30] This figure shows an example configuration of an information processing system according to Embodiment 10. [Figure 31] This flowchart shows an example of information processing according to Embodiment 10. [Figure 32] This flowchart shows an example of the network policy determination process according to Embodiment 10. [Figure 33] This figure shows an example configuration of the information processing system according to Embodiment 11. [Figure 34] This figure shows an example of the functional configuration of the learning policy determination unit according to Embodiment 11. [Figure 35] This flowchart shows an example of the second policy decision model learning process according to Embodiment 11. [Figure 36] This figure shows an example configuration of an information processing system according to Embodiment 12. [Figure 37] This flowchart shows an example of information processing according to Embodiment 12. [Figure 38] This flowchart shows an example of the network policy determination process according to Embodiment 12. [Figure 39] This figure shows an example configuration of an information processing system according to Embodiment 13. [Figure 40] This figure shows an example of the functional configuration of the learning policy determination unit according to Embodiment 13. [Figure 41] This flowchart shows an example of the third policy decision model learning process according to Embodiment 13. [Modes for carrying out the invention]

[0011] Hereinafter, one embodiment of the present invention will be described with reference to the drawings. In all drawings, similar components are denoted by the same reference numerals, and their descriptions are omitted where appropriate.

[0012] <Embodiment 1> (overview) Figure 1 is a diagram showing an overview of the information processing system 100 according to Embodiment 1. The information processing system 100 includes a quality estimation unit 111 and an image processing unit 113.

[0013] The quality estimation unit 111 estimates the quality information of the target image. The image processing unit 113 processes the target image using a neural network that corresponds to a network policy determined based on the quality information of the target image.

[0014] This information processing system 100 makes it possible to perform authentication regardless of the quality of the target image while suppressing an increase in processing time.

[0015] Figure 2 shows an overview of the information processing device 102 according to Embodiment 1. The information processing device 102 includes a quality estimation unit 111 and an image processing unit 113.

[0016] The quality estimation unit 111 estimates the quality information of the target image. The image processing unit 113 processes the target image using a neural network that corresponds to a network policy determined based on the quality information of the target image.

[0017] This information processing device 102 makes it possible to perform authentication regardless of the quality of the target image while suppressing an increase in processing time.

[0018] Figure 3 is a flowchart showing an overview of the information processing according to Embodiment 1.

[0019] The quality estimation unit 111 estimates the quality information of the target image (step S103).

[0020] The image processing unit 113 processes the target image using a neural network that corresponds to a network policy determined based on the quality information of the target image (step S105).

[0021] This information processing method makes it possible to perform authentication regardless of the quality of the target image while keeping processing time down.

[0022] The following describes a detailed example of the information processing system 100 according to Embodiment 1.

[0023] (detail) The technology described in Patent Document 1 above uses so-called image super-resolution technology to improve resolution. However, image super-resolution has the problem of requiring a long processing time.

[0024] One example of the purpose of this disclosure is, in light of these circumstances, to provide an information processing system, information processing device, information processing method, and recording medium that solve the problem of performing authentication regardless of the quality of the target image while suppressing an increase in processing time.

[0025] (Example of the configuration of the information processing system 100 according to Embodiment 1) Figure 4 shows an example of the configuration of the information processing system 100 according to Embodiment 1. The information processing system 100 is a system for performing processing using an image of an object that has been photographed. In this embodiment, an example is given in which the processing using the image of the object is an authentication process for authenticating the object.

[0026] The information processing system 100 comprises an imaging device 101 and an information processing device 102.

[0027] The imaging device 101 and the information processing device 102 are connected via a network configured by wire, wireless, or a combination thereof, and they send and receive information from each other via the network.

[0028] The subject in this embodiment is a human being. However, the subject is not limited to a human being; it may also be an animal, such as a dog or a snake.

[0029] Furthermore, the authentication process according to this embodiment performs iris authentication. In this embodiment, an example of iris authentication using an image of the right eye iris will be described. However, the authentication process is not limited to this, and any image-based authentication process is acceptable, such as iris authentication using an image of the left eye iris or both eyes' irises, facial authentication using a facial image showing the entire face, or vein authentication using an image showing veins.

[0030] The imaging device 101 photographs the target and generates an image of the target.

[0031] The target image is an image that includes the target region used for authentication processing. In this embodiment, as described above, iris authentication is performed using an image, so the target region is the iris region in which the iris is reflected. More specifically, in this embodiment, iris authentication is performed using an image of the right eye iris, so the target region is the iris region in which the right eye iris is reflected. However, iris authentication may also be performed using an image of the left eye iris, in which case the target region may be the iris region in which the right eye iris is reflected.

[0032] (Example of the functional configuration of the information processing device 102 according to Embodiment 1) The information processing device 102 according to this embodiment uses the target image generated by the imaging device 101 to authenticate the object that appears in the target image.

[0033] In this embodiment, as will be described later, an example is used in which the information processing device 102 is equipped with an authentication unit 114 that performs authentication processing (authentication processing) for the target. However, the information processing device 102 only needs to perform processing to authenticate the target using the target image, and the authentication processing may be performed by another information processing device, for example, which is not shown. In this case, it is preferable that the other information processing device is equipped with the authentication unit 114.

[0034] As shown in Figure 4, the information processing device 102 functionally comprises, for example, a quality estimation unit 111, a policy determination unit 112, an image processing unit 113, and an authentication unit 114.

[0035] The quality estimation unit 111 estimates the quality information of the target image.

[0036] Quality information is information that indicates the quality of the target image. In this embodiment, quality is the resolution of the target image, for example, the resolution of the target region included in the target image. More specifically, in this embodiment, quality is the resolution of the iris region in which the right eye iris used in the authentication process is reflected. When quality is resolution, the lower the resolution, the lower the quality.

[0037] The quality estimation unit 111 may further estimate the position of the target region in the target image. The position of the target region is information for identifying the position of the target region in the target image. In this embodiment, since the target region is the iris region described above, the position of the target region in this embodiment is the position of the right eye iris in the target image. Hereinafter, the position of the right eye iris in the target image will also be referred to as the "eye position".

[0038] Eye position includes, for example, the position of the pupil center, pupil diameter, and iris diameter. Pupil diameter is the diameter or radius of the pupil. Iris diameter is the diameter or radius of the outer edge of the iris. By using such eye position, the position of the right eye iris in the target image can be determined.

[0039] Furthermore, the quality information indicates quality that is not limited to the resolution of the target image, but may include one or more of the following: the resolution of the target image, the iris diameter of the iris reflected in the target image, the estimated distance between the imaging device 101 and the target, and the presence or degree of blur. This estimated distance may be estimated from information used for focusing on the imaging device 101, or it may be estimated based on information from a distance measuring sensor (not shown) for estimating the distance to the target. In addition, the resolution of the target image is not limited to the target area, but may include, for example, the resolution of a predetermined area included in the target image.

[0040] Furthermore, quality information may be represented, for example, as a vector quantity. When quality information is represented by multiple parameters, using a vector quantity to represent the quality information allows the multiple parameters to be treated as a single variable.

[0041] The policy determination unit 112 determines a network policy based on the quality information estimated by the quality estimation unit 111. This network policy is applied to the neural network used for image processing by the image processing unit 113, which will be described later.

[0042] The policy determination unit 112 may, for example, determine a network policy that satisfies at least one of the following conditions (A) and (B), regardless of the quality information of the target image.

[0043] (A) The processing time using the neural network must be within a predetermined range. (B) The accuracy of authentication using the target image must be within a predetermined range.

[0044] The image processing unit 113 processes the target image using a neural network corresponding to the network policy determined by the policy determination unit 112. In processing the target image according to this embodiment, the image processing unit 113 extracts features from the target image, for example.

[0045] Here, the image processing unit 113 may process the target image using a neural network that performs calculations using more detailed information the lower the quality of the target image (i.e., the quality indicated by the quality information), depending on the network policy. In other words, the policy determination unit 112 determines a network policy that performs calculations using more detailed information the lower the quality of the target image.

[0046] Figure 5 shows an example configuration of the neural network NN1 according to Embodiment 1. The neural network NN1 is used for processing the target image performed by the image processing unit 113. Here, an example in which the neural network NN1 is a convolutional neural network is used for explanation. However, the neural network NN1 is not limited to this, and for example, a vision transformer may be used.

[0047] The neural network NN1, for example, includes multiple layers. The multiple layers included in the neural network NN1 consist of a pre-stage group FNN1 that includes an input layer INL, and a post-stage group RNN that includes a final layer OUTL.

[0048] The number of layers in the preceding group FNN1 and the succeeding group RNN may be changed as appropriate, but for example, the number of layers in the preceding group FNN1 may be greater than the number of layers in the succeeding group RNN. In other words, the preceding group FNN1 may be located on the input side of the layer that is located in the middle of the multiple layers that make up the neural network NN1.

[0049] The preceding group FNN1 includes, for example, an input layer INL, at least one convolutional layer FCL, and an intermediate output layer MOUT. The preceding group FNN1 may also include batch norm layers, activation functions, etc., similar to a typical convolutional neural network (not shown).

[0050] The input layer INL acquires the input image for the neural network NN1 and propagates the input image to subsequent layers, for example. The input image is the target image. The image processing unit 113 may also extract an image contained within the target image (for example, an image of the target region) and use this extracted image as the input image. In other words, the input image can be any image obtained from the target image.

[0051] Each layer, including the convolutional layer FCL, processes the input image, for example, by extracting features from the input image. The intermediate output layer MOUT outputs the processing results from the neural network NN1 as intermediate features. Note that the intermediate output layer MOUT may also be a convolutional layer.

[0052] Here, the network policy determined by the policy determination unit 112 is applied to image processing using the preceding group FNN1. In other words, the policy determination unit 112 determines the network policy to be applied to the preceding group FNN1. Specifically, for example, the policy determination unit 112 determines a network policy that uses more detailed information for calculations the lower the quality of the target image.

[0053] In this case, the policy determination unit 112 may determine a network policy that satisfies at least one of (A) and (B) above. For example, the policy determination unit 112 may pre-store policy information that defines network policies according to the quality of the target image, and determine a network policy according to the quality of the target image by obtaining a network policy according to the quality of the target image from the policy information. In the policy information, for example, it is preferable that a network policy is pre-defined in which the preceding group FNN1 performs calculations using detailed information according to the quality of the target image, so as to satisfy at least one of (A) and (B) above.

[0054] The subsequent group RNN includes, for example, an intermediate input layer MIN, at least one convolutional layer RCL, and a final layer (output layer) OUTL. The subsequent group RNN may also include batch norm layers, activation functions, etc., similar to a typical convolutional neural network (not shown).

[0055] The intermediate input layer MIN acquires intermediate features from the preceding group FNN1. Each layer, including the convolutional layer RCL, further processes the intermediate features, for example, by extracting features from the input image. Note that the intermediate input layer MIN may also be a convolutional layer.

[0056] The final layer OUTL outputs the final processing results of the neural network NN1. The final layer OUTL outputs feature vectors based on the results of further processing in each layer, including, for example, the convolutional layer RCL.

[0057] Thus, the lower the quality of the target image, the more detailed the information used in the calculations, which helps to minimize the decrease in authentication accuracy even with low-quality images. On the other hand, because detailed information is used in the calculations, processing time using neural networks increases for low-quality images compared to high-quality images.

[0058] As described above, for example, the policy determination unit 112 may determine a network policy such that the processing time using the neural network (A) is within a predetermined range, regardless of the quality information of the target image (i.e., the quality indicated by the quality information). This suppresses the increase in processing time (processing time using the neural network) when the target image is of low quality, and makes it possible to keep the processing time using the neural network roughly the same regardless of the quality of the target image.

[0059] Alternatively, for example, the policy determination unit 112 may determine a network policy such that (B) the accuracy of authentication using the target image is within a predetermined range, regardless of the quality information of the target image (i.e., the quality indicated by the quality information). This suppresses the decrease in authentication accuracy when the target image is of low quality, and makes it possible to maintain a similar level of authentication accuracy in the authentication process regardless of the quality of the target image.

[0060] The authentication unit 114 performs authentication processing using the results of the image processing unit 113's processing of the target image. Authentication processing is the process of authenticating the target using the results of the processing of the target image.

[0061] For example, when the authentication unit 114 obtains the feature quantities extracted by the image processing unit 113, it performs authentication based on whether or not the obtained feature quantities are registered in the registered data. Although not shown in the diagram, the authentication unit 114 may store the registered data in advance.

[0062] Up to this point, we have mainly described examples of the functional configuration of the information processing system 100. From here, we will describe examples of the physical configuration of the information processing system 100.

[0063] (Physical configuration of information processing system 100) The information processing system 100 according to this embodiment is physically composed of, for example, a shooting device 101 and an information processing device 102. The shooting device 101 is, for example, a camera such as a near-infrared camera.

[0064] Figure 6 shows an example of the physical configuration of the information processing device 102 according to Embodiment 1. The information processing device 102 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.

[0065] Bus 1010 is a data transmission path for the processor 1020, memory 1030, storage device 1040, network interface 1050, input interface 1060, and output interface 1070 to send and receive data to and from each other. However, the method of connecting the processor 1020 and other components to each other is not limited to bus connection.

[0066] Processor 1020 is a processor implemented in components such as the CPU (Central Processing Unit) and GPU (Graphics Processing Unit).

[0067] Memory 1030 is a main memory device implemented using RAM (Random Access Memory), etc.

[0068] The storage device 1040 is an auxiliary storage device implemented as an HDD (Hard Disk Drive), SSD (Solid State Drive), memory card, or ROM (Read Only Memory). The storage device 1040 stores program modules for realizing the functions of the information processing device 102. The processor 1020 loads each of these program modules into memory 1030 and executes them, thereby realizing the functions corresponding to those program modules.

[0069] The network interface 1050 is an interface for connecting the information processing device 102 to a network.

[0070] The input interface 1060 is an interface for the user to input information, and consists of, for example, a touch panel, a keyboard, a mouse, etc.

[0071] The output interface 1070 is an interface for presenting information to the user and consists of, for example, an LCD panel or an OLED (Electro-Luminescence) panel.

[0072] The functions of the information processing device 102 may also be implemented using multiple physical information processing devices. In this case, each information processing device may be physically configured similarly to the information processing device 102 shown in Figure 6. Furthermore, the multiple information processing devices may be connected to each other via a network configured by wire, wireless, or a combination thereof, and configured to send and receive information from each other via the network.

[0073] We have now described an example configuration of the information processing system 100 according to Embodiment 1. From here, we will describe an example of operation of the information processing system 100 according to Embodiment 1.

[0074] (Example of operation of the information processing system 100 according to Embodiment 1) The information processing system 100 performs information processing. The information processing according to this embodiment is a process that uses an image of the subject taken to authenticate the subject that appears in the image.

[0075] In this embodiment, an example is described using an authentication process that performs authentication of the target, but the information processing only needs to be a process for authenticating the target using the target image, and does not necessarily have to include an authentication process.

[0076] Information processing may be initiated, for example, by receiving a signal from a sensor (not shown). This sensor is, for example, one that detects the presence or entry of a person into the area being photographed by the imaging device 101. However, the method for initiating information processing is not limited to this.

[0077] Figure 7 is a flowchart showing an example of information processing according to Embodiment 1.

[0078] The imaging device 101 photographs the target (step S101). As a result, the imaging device 101 generates an image of the target.

[0079] Figure 8 shows an example of a target image according to Embodiment 1. The target image according to this embodiment is an eye image including both eyes of the target. The eye image is an example of an image including the target region according to this embodiment. As described above, the target region according to this embodiment is the iris region in which the iris of the right eye is reflected.

[0080] The target image may be modified in various ways depending on the type of authentication performed using it. For example, in the case of facial recognition, it may be a facial image including the target's face and background, or an image including the target's upper body or entire body. In this case, the target area is the facial area in which the face is visible. in the region Yes, for example, when vein authentication is performed, The target area is, Area where veins are visible Includes A blank image is also acceptable.

[0081] Furthermore, for example, the target image used for iris authentication is not limited to an eye image, but may also be a face image, an image including the upper body or the entire body of the subject, etc. Moreover, the eye image is not limited to an eye image including both eyes of the subject, but may also be a single-eye image including one of the predetermined left or right eyes of the subject. The eye image illustrated in Figure 8 includes not only the iris but also the pupil, white of the eye, eyelid, eyelashes, etc., but the eye image only needs to include at least the iris region used for authentication processing, and one or more of the pupil, white of the eye, eyelid, eyelashes, etc. may not be included.

[0082] Refer to Figure 7 again. The quality estimation unit 111 estimates the location of the target region in the target image (step S102).

[0083] The quality estimation unit 111 according to this embodiment estimates, for example, the eye position including the position of the pupil center of the right eye, the pupil diameter, and the iris diameter from the target image.

[0084] The quality estimation unit 111 may use general techniques such as pattern matching or machine learning models to estimate eye positions. When using a machine learning model, the quality estimation unit 111 may use a trained model that has been trained to estimate eye positions from target images, and use the target image as input to estimate the eye positions in the target image. In training this trained model, supervised learning may be performed using training target images as input and training data that includes the eye positions in the target images.

[0085] The quality estimation unit 111 estimates the quality of the target image (step S103). As a result, the quality estimation unit 111 generates quality information including the quality of the target image.

[0086] The quality estimation unit 111 according to this embodiment identifies the iris region (i.e., the region in which the iris of the target's right eye is reflected) included in the target image, for example, based on the eye position estimated in step S102. The iris region is generally an annular region surrounded by concentric circles with diameters corresponding to the pupil diameter and iris diameter, respectively, centered on the pupil center. Note that, as shown in Figure 8, a part of the iris region may be obscured by the eyelid, etc. but When covered, the iris region will have a shape where the part covered by the eyelid or other tissue is missing from the ring-shaped area.

[0087] The quality estimation unit 111 determines the resolution of the identified iris region. Resolution is, for example, the image size. In this case, the quality estimation unit 111 determines the image size of the identified iris region. The image size of the iris region is the number of pixels in the iris region.

[0088] Alternatively, the quality estimation unit 111 may determine the quality of the target image using a learning model. In this case, the learning model should be one that has been trained to estimate the quality of the target image, and should take the target image as input to estimate the quality of the target image. Furthermore, in the training of this learning model, supervised learning should be performed using training target images as input and training data that includes the quality of the target image.

[0089] The iris diameter may be used as the image size. In this case, step S102 will include a process for estimating the quality of the target image. Even if a value other than the iris diameter is used for the quality of the target image, a common learning model may be used in steps S102 and S103. In this case, the learning model should be one that has been trained to estimate eye position and the quality of the target image from the target image, and should take the target image as input to estimate the eye position and the quality of the target image in the target image. Furthermore, in training this learning model, supervised learning should be performed using training target images as input and training data including eye position and the quality of the target image from the target image.

[0090] The policy determination unit 112 determines the network policy based on the quality of the target image estimated in step S103 (step S104).

[0091] In detail, for example, the policy determination unit 112 may determine a network policy such that at least one of (A) and (B) above is satisfied, regardless of the quality information of the target image.

[0092] The image processing unit 113 processes the target image using the neural network NN1 according to the network policy determined in step S104 (step S105).

[0093] In detail, for example, the image processing unit 113 uses a neural network NN1 according to the network policy to extract feature quantities from the iris region estimated in step S102.

[0094] Here, for example, suppose that in step S104, the policy determination unit 112 determines a network policy that satisfies both (A) and (B) regardless of the quality information of the target image. This makes it possible to make both the processing time using the neural network NN1 (i.e., processing for extracting features, for example) and the accuracy of authentication performed in the authentication process roughly the same, regardless of the quality of the target image.

[0095] The authentication unit 114 performs authentication processing using the result of the processing in step S105 (step S106).

[0096] In detail, for example, the authentication unit 114 acquires the feature quantities extracted by the image processing unit 113. The authentication unit 114 compares the acquired feature quantities with the feature quantities obtained from pre-stored registered data, and authenticates the target based on the result of the comparison.

[0097] More specifically, the authentication unit 114 calculates the similarity between the acquired features and the features obtained from the registered data. The authentication unit 114 authenticates the target based on the result of comparing the similarity with a predetermined threshold. For example, the authentication unit 114 determines that the authentication of the target was successful if the similarity is equal to or greater than the threshold. Alternatively, the authentication unit 114 determines that the authentication of the target failed if the similarity is less than the threshold. Note that the authentication method is not limited to this, and a general method may be used.

[0098] (Effects / Actions) As described above, according to this embodiment, the information processing system 100 comprises a quality estimation unit 111 and an image processing unit 113.

[0099] The quality estimation unit 111 estimates the quality information of the target image. The image processing unit 113 processes the target image using a neural network NN1 that corresponds to a network policy determined based on the quality information of the target image.

[0100] This allows the processing in the neural network NN1 to be changed according to the quality of the target image. Therefore, it becomes possible to perform authentication regardless of the quality of the target image while keeping the increase in processing time to a minimum.

[0101] According to this embodiment, the image processing unit 113 extracts features from the target image using a neural network NN1 according to the network policy.

[0102] This allows authentication processing to be performed using features. The process for extracting features is carried out using a neural network NN1 according to the network policy. Therefore, it becomes possible to perform authentication regardless of the quality of the target image while keeping the increase in processing time to a minimum.

[0103] According to this embodiment, the target image is an eye image that includes the iris region in which the iris is reflected. The quality information includes the iris diameter.

[0104] This allows the processing in the neural network NN1 to be changed according to the iris diameter of the target image. Therefore, it becomes possible to perform authentication such as iris recognition regardless of the quality of the target image while suppressing the increase in processing time.

[0105] According to this embodiment, the quality information includes the resolution of the target image.

[0106] This allows the processing in the neural network NN1 to be changed according to the resolution of the target image. Therefore, it becomes possible to perform authentication regardless of the quality of the target image while suppressing the increase in processing time.

[0107] According to this embodiment, the information processing system 100 further includes a policy determination unit 112 that determines a network policy based on quality information.

[0108] This allows for the determination of a network policy based on the quality of the target image. The processing in the neural network NN1 can then be modified according to this determined network policy. Therefore, authentication can be performed regardless of the quality of the target image while minimizing the increase in processing time.

[0109] According to this embodiment, the policy determination unit 112 determines a network policy such that the processing time using the neural network NN1 falls within a predetermined range, regardless of the resolution of the target image.

[0110] This allows the processing in the neural network NN1 to be modified according to the resolution of the target image, while suppressing the increase in processing time that results from such modifications. Therefore, it becomes possible to perform authentication regardless of the quality of the target image while suppressing the increase in processing time.

[0111] According to this embodiment, the neural network NN1 consists of a pre-stage group FNN2 including an input layer INL and a post-stage group RNN including a final layer OUTL. The network policy is applied to image processing using the pre-stage group FNN2.

[0112] This allows the processing in the neural network NN1 to be changed according to the resolution of the target image. Furthermore, it can suppress the increase in processing time compared to improving the quality of the input image fed into the neural network NN1 using image super-resolution technology. Therefore, it becomes possible to perform authentication regardless of the quality of the target image while suppressing the increase in processing time.

[0113] <Embodiment 2> This embodiment describes detailed examples of the network policy and the processing performed by the image processing unit 113. In this embodiment, in order to simplify the explanation, explanations that overlap with the embodiments described above will be omitted as appropriate.

[0114] The information processing system according to this embodiment is preferably configured in a manner similar to the information processing system 100 according to Embodiment 1 illustrated in Figures 4 and 6. That is, for example, the image processing unit 113 processes the target image using a neural network according to the network policy, similar to Embodiment 1. Also, for example, the neural network constituting the image processing unit 113 performs calculations using more detailed information the lower the quality of the target image, according to the network policy, similar to Embodiment 1.

[0115] The image processing unit 113 according to this embodiment converts the iris region included in the target image into a standardized image of a predetermined shape, and processes the converted standardized image as an input image to the neural network.

[0116] Figure 9 shows an example of the functional configuration of the image processing unit 113 according to Embodiment 2. The image processing unit 113 includes a first processing unit 113a and a second processing unit 113b.

[0117] The first processing unit 113a converts a target region (i.e., the iris region) included in the target image into a standard image of a predetermined shape. Specifically, for example, the first processing unit 113a extracts the iris region from the target image and converts the extracted iris region into a standard image of a predetermined shape. In this way, the first processing unit 113a generates a standard image based on the target image.

[0118] The second processing unit 113b processes the standard image acquired by the first processing unit 113a using a neural network corresponding to the network policy determined by the policy determination unit 112. In processing the standard image according to this embodiment, the second processing unit 113b extracts features from the standard image, for example.

[0119] Figure 10 shows an example configuration of the neural network NN2 according to Embodiment 2. The neural network NN2 is used for processing standard images performed by the second processing unit 113b according to Embodiment 2. In this embodiment as well, the explanation will use an example in which the neural network NN2 is a convolutional neural network.

[0120] The neural network NN2 includes, for example, multiple layers, similar to Embodiment 1. The multiple layers included in the neural network NN2 consist of a pre-stage group FNN2 including an input layer INL, and a post-stage group RNN similar to Embodiment 1.

[0121] The preceding group FNN2 includes an input layer INL, CLN convolutional layers FCL, and an intermediate output layer MOUT.

[0122] The input layer INL acquires the input image, similar to Embodiment 1, and propagates the input image to subsequent layers, for example. The input image according to this embodiment is a standard image generated by the first processing unit 113a.

[0123] Each layer, including CLN convolutional layers FCL, processes the input image, for example, by extracting features from the input image. The intermediate output layer MOUT outputs the processing results from the neural network NN2 as intermediate features. Note that the intermediate output layer MOUT may also be a convolutional layer.

[0124] In this embodiment as well, the network policy determined by the policy determination unit 112 is applied to image processing using the preceding group FNN2.

[0125] In detail, for example, the policy determination unit 112 according to this embodiment determines a network policy such that the number of convolutional layers FCLs (CLN) included in the preceding group FNN2 increases as the resolution of the target image decreases. This network policy is an example of a network policy in which the preceding group FNN2 performs calculations using more detailed information as the quality of the target image decreases.

[0126] The information processing according to this embodiment is preferably configured in general the same way as the information processing according to Embodiment 1 illustrated in Figure 7. That is, for example, in step S105, the image processing unit 113 processes the target image using a neural network according to a network policy determined based on the quality information of the target image.

[0127] Figure 11 is a flowchart showing a detailed example of the image processing (step S105) according to Embodiment 2. In the image processing (step S105), as explained with reference to Figure 7, the image processing unit 113 may process the target image using a neural network according to the network policy determined in step S104.

[0128] The first processing unit 113a extracts the target region from the target image (step S105a).

[0129] In detail, for example, the target region is the iris region. Therefore, the first processing unit 113a extracts the iris region in a roughly annular shape.

[0130] Figure 12 shows an example of the iris region according to Embodiment 2. In this figure, the iris region is a hatched area and is surrounded by inner and outer circumferences that are concentric circles centered on the pupil center O. The inner circumference is the circumference corresponding to the pupil diameter. The outer circumference is the circumference corresponding to the iris diameter. Such an iris region is, for example, defined with the pupil center O as the origin, and a radial length R and a predetermined reference direction versus It can be represented in polar coordinates using the angle θ. Note that the area inside the inner circumference corresponds to the pupil, but this area does not need to be included in the iris region.

[0131] Refer to Figure 11 again. The first processing unit 113a converts the target region extracted in step S105a into a standard image of a predetermined shape (step S105b). In this way, the first processing unit 113a generates a standard image.

[0132] Figure 13 shows an example of a standard image according to Embodiment 2. In this figure, an example is shown where the predetermined shape is a rectangle, but the predetermined shape is not limited to this.

[0133] When converting the annular iris region illustrated in Figure 12 to the rectangular standard image illustrated in Figure 13, for example, the height H of the standard image corresponds to the width R1 of the annular iris region. Also, for example, the width W of the standard image corresponds to the outer and inner sides of the annular iris region.

[0134] In step S105b, the first processing unit 113a may convert the iris region into a standard image of a size corresponding to its quality information (e.g., resolution). That is, the size of the standard image may vary depending on the quality of the target image (e.g., the image size of the iris region). In this case, the first processing unit 113a may convert the target region into a larger standard image the higher the quality of the target image (e.g., the larger the image size of the iris region). Also, if the standard image is rectangular, its aspect ratio may remain constant.

[0135] Any method may be used to convert the ring-shaped iris region into a rectangular standard image.

[0136] For example, the first processing unit 113a may convert the target region to standard images of discretely different sizes depending on the quality of the target image. In this case, for example, the first processing unit 113a may pre-store conversion pattern data indicating the sizes of multiple standard images corresponding to the quality of the target image. Then, the first processing unit 113a may obtain the size of the standard image corresponding to the quality of the target region estimated in step S103 from the conversion pattern data, and convert the extracted target region in step S105a to a standard image of the obtained size. Such conversions may be carried out using appropriate methods such as conversion from polar coordinates to Cartesian coordinates, or conversions to enlarge and reduce the image.

[0137] Alternatively, for example, the first processing unit 113a may convert the target area into standard images of continuously different sizes depending on the quality of the target image. In this case, for example, the first processing unit 113a may convert the iris region extracted in step S105a into a rectangular image using a conversion from a polar coordinate system to a Cartesian coordinate system. If the aspect ratio of this rectangular image differs from that of a predetermined standard image, the first processing unit 113a may convert the rectangular image into a standard image with a predetermined aspect ratio using at least one of an image enlargement or reduction conversion.

[0138] In this case, to convert the aspect ratio of a rectangular image to a predetermined aspect ratio of a standard image, the extent to which the vertical and horizontal directions are converted should be determined according to predetermined conversion conditions.

[0139] For example, if the conversion condition is not to convert the horizontal direction, the first processing unit 113a may convert the rectangular image into a standard image which is a rectangle with a predetermined aspect ratio by enlarging or reducing it in the vertical direction. Also, for example, if the conversion condition is not to convert the vertical direction, the first processing unit 113a may convert the rectangular image into a standard image which is a rectangle with a predetermined aspect ratio by enlarging or reducing it in the horizontal direction. Note that the conversion conditions are not limited to these.

[0140] By performing this conversion, the iris region can be transformed into a standardized image of a size corresponding to its quality information (e.g., resolution). Furthermore, the standardized image will have a quality (e.g., resolution) that corresponds to the quality information of the iris region.

[0141] Refer to Figure 11 again. The second processing unit 113b processes the standard image generated in step S105b using the neural network NN2 according to the network policy determined in step S104 (step S105c). Then, the second processing unit 113b returns to the information processing shown in Figure 7.

[0142] In detail, for example, the size of a standard image varies discretely or continuously depending on the quality of the target image, as mentioned above. Examples of network policies are described for each case.

[0143] (Example of a network policy when the sizes of standard images vary discretely) For example, let's assume that the quality of the target image is the resolution of the input image. Also, if the resolution of the input image is a predetermined reference resolution, let's assume that the number of convolutional layers FCL used in the calculations of the preceding group FNN2, CLN, is the reference value RFN1.

[0144] In this case, the policy determination unit 112 may determine a network policy in which, when the resolution of the input image becomes 1 / N of the reference resolution, the number of convolutional layers FCLs (CLN) used in the calculations of the preceding group FNN is increased to RFN1*N*N. Each of N and RFN1 is an integer greater than or equal to 1. The "*" represents multiplication.

[0145] As a result, the second processing unit 113b processes the input image using a pre-stage group FNN2 that includes several CLN convolutional layers FCL according to the network policy, and outputs intermediate features.

[0146] Note that the network policy is not limited to this, and may be modified as appropriate, for example, to determine how detailed the information used for calculations depends on the quality of the target image.

[0147] (Example of a network policy when the size of standard images varies continuously) For example, let's assume that the quality of the target image is the resolution of the input image, similar to the case where the sizes are discretely different. Also, let's assume that the number of convolutional layers FCLs (CLN) used in the calculations of the preceding group FNN2 is the reference value RFN1, given that the resolution of the input image is a predetermined reference resolution.

[0148] In this case, the policy determination unit 112 may determine a network policy in which, when the resolution of the input image becomes 1 / α of the reference resolution, the number of convolutional layers FCLs (CLN) used in the calculations of the preceding FNN group is increased to [RFN1*α*α]. Here, α is a real number, for example, satisfying 0 < α ≤ 1. Also, the symbol [X] is a Gaussian function (also called the floor function) and represents the largest integer not exceeding the real number X in the symbol. That is, [RFN1*α*α] represents the largest integer not exceeding RFN1*α*α.

[0149] As a result, the second processing unit 113b processes the input image using a pre-stage group FNN2 that includes several CLN convolutional layers FCL according to the network policy, and outputs intermediate features. Note that the network policy is not limited to this, and the degree to which detailed information is used for calculations may be changed as appropriate, for example, depending on the quality of the target image, just as when the sizes of standard images are discretely different.

[0150] Furthermore, if the sizes of the standard images vary continuously, the FNN2 group should adjust the shape of the intermediate feature map so that the intermediate features are of a shape (i.e., size) that can be input to the RNN group.

[0151] For example, interpolation can be used to adjust the shape of the intermediate feature map. For instance, suppose a 3x3 filter is used to perform a convolution with a slide S and padding PD, and the shape of the input feature map is (Hin, Win). If the shape of the output feature map from this convolution is (Hout, Wout), then the input feature map should be adjusted using interpolation so that it satisfies the relationships Hin = S*(Hout-1) + 3-2*PD and Win = S*(out-1) + 3-2*PD. Suitable interpolation methods include the nearest neighbor method, linear interpolation, bilinear interpolation, and bicubic interpolation. Such interpolation should ideally be performed before the first convolutional layer FCL (i.e., the convolutional layer FCL closest to the input) in the preceding group FNN2.

[0152] For example, the value of the dilation parameter used in the convolutional layer FCL can be used to adjust the map shape of intermediate features. The dilation parameter is used to adjust the interval between values ​​on the map obtained by multiplying the convolutional layer FCL with the filter.

[0153] Furthermore, these are not the only methods for adjusting the map shape of intermediate features.

[0154] (Effects / Actions) As described above, according to this embodiment, the image processing unit 113 includes a first processing unit 113a and a second processing unit 113b.

[0155] The first processing unit 113a converts the iris region included in the target image into a standardized image of a predetermined shape. The second processing unit 113b processes the standardized image using a neural network NN2 according to the network policy. The first processing unit 113a converts the iris region into a standardized image of a size according to the quality information.

[0156] This allows the shape of the input image to the neural network NN2 to be normalized to a predetermined shape, making processing by the neural network NN2 easier than when the shape of the input image is undefined.

[0157] Furthermore, since an image corresponding to the iris region can be used as the input image, the feature quantities of the iris region can be extracted with high accuracy, thereby improving the accuracy of iris recognition.

[0158] Furthermore, generally, when image resolution is improved using image super-resolution technology, the neural network processes the high-resolution image, which often increases the processing time of the neural network. In this embodiment, since the standard image is sized according to its resolution, the increase in processing time of the neural network NN2 can be suppressed compared to processing high-resolution images uniformly with the neural network NN2.

[0159] Therefore, it becomes possible to perform authentication with high accuracy regardless of the quality of the target image, while keeping the increase in processing time to a minimum.

[0160] According to this embodiment, the neural network NN2 includes multiple convolutional layers FCL and RCL. The policy determination unit 112 determines the network policy such that the number of convolutional layers FCL included in the preceding group FNN2 (CLN) increases as the resolution of the target image decreases.

[0161] This allows for the use of more convolutional layers (FCLs) when the target image quality is low, enabling calculations using more detailed information. Therefore, the neural network NN2 can process input images to enable accurate authentication regardless of the input image quality. Furthermore, this approach minimizes the increase in processing time compared to improving the quality of the input image fed into the neural network NN2 using image super-resolution techniques.

[0162] Therefore, it becomes possible to perform authentication regardless of the quality of the target image while keeping the increase in processing time to a minimum.

[0163] <Embodiment 3> Embodiment 2 describes an example in which the number of convolutional layers (CLN) is increased as the resolution of the target image decreases, thereby enabling calculations to be performed using more detailed information as the quality of the target image decreases. Embodiment 3 describes another example in which calculations are performed using more detailed information as the quality of the target image decreases, by increasing the number of channels in the calculations of the convolutional layer (FCL) as the resolution of the target image decreases. In this embodiment, in order to simplify the explanation, explanations that overlap with the embodiments described above will be omitted as appropriate.

[0164] In this embodiment, the configuration of the neural network used for processing standard images by the second processing unit 113b differs from that of Embodiment 2. Aside from this, the information processing system according to this embodiment may be configured in general the same way as the information processing system according to Embodiment 2.

[0165] Figure 14 shows an example configuration of the neural network NN3 according to Embodiment 3. The neural network NN3 is used for processing standard images performed by the second processing unit 113b according to Embodiment 3. In this embodiment as well, the explanation will use an example in which the neural network NN3 is a convolutional neural network.

[0166] The neural network NN3 includes, for example, multiple layers, similar to Embodiment 1. The multiple layers included in the neural network NN3 consist of a pre-stage group FNN2 including an input layer INL, and a post-stage group RNN similar to Embodiment 1.

[0167] The preceding group FNN3 includes an input layer INL similar to that in Embodiment 1, P convolutional layers FCL_1 to FCL_P, and an intermediate output layer MOUT, where P is an integer greater than or equal to 2.

[0168] The convolutional layers FCL_1 to FCL_P have different numbers of channels. The number of channels corresponds, for example, to the number of filters used in each of the convolutional layers FCL_1 to FCL_P. Each layer containing any one of the convolutional layers FCL_1 to FCL_P processes the input image and, for example, extracts features from the input image. In this embodiment, we will explain using an example where there is one convolutional layer FCL with the same number of channels, but the preceding group FNN3 may contain multiple convolutional layers FCL with the same number of channels.

[0169] The intermediate output layer MOUT outputs the processing results from the neural network NN2 as intermediate features.

[0170] In this embodiment, since the number of channels in the convolutional layers FCL_1 to FCL_P are different, the shape of the feature maps output from each of the convolutional layers FCL_1 to FCL_P is different. Regardless of which of the convolutional layers FCL_1 to FCL_P is used, the intermediate output layer MOUT may adjust the size of the intermediate features (the shape of the feature map representing the intermediate features) so that the intermediate features can be input to the subsequent group RNN. For example, a 1x1 convolution may be used for this adjustment.

[0171] In this embodiment as well, the network policy determined by the policy determination unit 112 is applied to image processing using the preceding group FNN3.

[0172] In detail, for example, the policy determination unit 112 according to this embodiment determines a network policy such that the lower the resolution of the target image, the more convolutional layer FCL with a larger number of channels is used in the preceding group FNN2. This network policy is an example of a network policy in which the preceding group FNN3 performs calculations using more detailed information as the quality of the target image decreases.

[0173] The information processing according to this embodiment may be configured in general the same way as the information processing according to Embodiment 1 illustrated in Figure 7. Furthermore, the details of the image processing (step S105) according to this embodiment may be configured in general the same way as the image processing (step S105) according to Embodiment 2 illustrated in Figure 11.

[0174] In other words, for example, in step S105c, the second processing unit 113b processes the standard image generated in step S105b using the neural network NN3 according to the network policy determined in step S104.

[0175] For example, suppose the sizes of the standard images are discretely different, and the quality of the target image is the resolution of the input image. Also, if the resolution of the input image is a predetermined reference resolution, suppose the number of channels in the convolutional layers FCL_1 to FCL_P used in the calculations of the preceding group FNN2 is the reference value RFN2. Here, RFN2 is an integer greater than or equal to 1.

[0176] In this case, the policy determination unit 112 may determine a network policy that increases the number of channels in the convolutional layers FCL_1 to FCL_P used for calculations in the preceding group FNN to RFN2*N*N when the resolution of the input image becomes 1 / N of the reference resolution.

[0177] As a result, the second processing unit 113b processes the input image using the pre-stage group FNN3, which includes a convolutional layer FCL with a number of channels according to the network policy, from among the convolutional layers FCL_1 to FCL_P, and outputs intermediate features. Note that the network policy is not limited to this, and may be changed as appropriate, for example, to what extent to which detailed information is used for calculations depending on the quality of the target image.

[0178] For example, if the size of the standard image is continuously different, and the resolution of the input image becomes 1 / α of the reference resolution, a network policy may be decided to increase the number of channels in the convolutional layers FCL_1 to FCL_P used in the calculations of the preceding FNN group to [RFN2*α*α]. The symbol [ ] here is the same Gaussian function as described above. That is, [RFN2*α*α] represents the largest integer not exceeding RFN2*α*α.

[0179] Furthermore, the network policy is not limited to this, and, for example, the level of detail used for calculations depending on the quality of the target image may be changed as appropriate, just as in the case where the sizes of standard images vary discretely.

[0180] Furthermore, when the sizes of the standard images differ continuously, the map shape of the intermediate features in the preceding group FNN3 should be adjusted, similar to Embodiment 2, so that the intermediate features can be input to the subsequent group RNN.

[0181] (Effects / Actions) As described above, according to this embodiment, the neural network NN3 includes multiple convolutional layers FCL_1 to FCL_P and RCL. The policy determination unit 112 determines the network policy such that the number of channels in the calculations performed by the convolutional layers FCL_1 to FCL_P included in the preceding group FNN3 increases as the resolution of the target image decreases.

[0182] Thus, the lower the quality of the target image, the more convolutional layers FCL_1 to FCL_P with a larger number of channels can be used, allowing for calculations using more detailed information. Therefore, the input image can be processed by the neural network NN3 to enable accurate authentication regardless of the quality of the input image. Furthermore, this approach minimizes the increase in processing time compared to improving the quality of the input image fed into the neural network NN3 using image super-resolution techniques.

[0183] Therefore, it becomes possible to perform authentication regardless of the quality of the target image while keeping the increase in processing time to a minimum.

[0184] <Embodiment 4> Embodiment 4 describes examples of training methods for neural networks NN1 to NN3. In this embodiment, the training method for neural network NN1 according to Embodiment 1 is used as an example. In addition, in this embodiment, in order to simplify the explanation, explanations that overlap with the embodiments described above are omitted as appropriate.

[0185] (Example of the configuration of the information processing system 400 according to Embodiment 4) Figure 15 shows an example configuration of the information processing system 400 according to Embodiment 4. The information processing system 400 includes an information processing device 403 in addition to the same imaging device 101 and information processing device 102 as in Embodiment 1.

[0186] The information processing device 403 is a device for training the neural network NN1. The information processing device 403 comprises a learning memory unit 421, a learning unit 422, and a learning policy determination unit 423.

[0187] The learning memory unit 421 is a memory unit for pre-storing learning data. The learning data includes the target image for training and the correct values. For example, when extracting features from a target image using a neural network NN1, the correct values ​​are the features of the target image for training.

[0188] The training image is simply an image used for training. Like the target image, the training image can be any image containing the target region, such as an eye image.

[0189] Furthermore, the learning memory unit 421 may pre-store learning data that includes multiple different learning target images. In this case, it is desirable that the multiple learning target images include learning target images of different qualities.

[0190] The learning unit 422 trains the neural network NN1 by inputting the training target images into the neural network NN1.

[0191] In detail, for example, the learning unit 422 includes a feature extraction unit 422a, a first loss acquisition unit 422b, and a first parameter update unit 422c.

[0192] The feature extraction unit 422a extracts feature quantities from the training target image using the neural network NN1 according to the network policy.

[0193] The first loss acquisition unit 422b calculates the first loss function by determining the error (first loss) between the extracted features and the correct values ​​of the training target images.

[0194] The first parameter update unit 422c calculates the gradient based on the first loss function and updates the parameters included in the neural network NN1.

[0195] The training policy determination unit 423 determines a network policy for the training target image based on the quality information and training probability distribution of the training target image.

[0196] In detail, for example, the learning policy determination unit 423 includes a first determination unit 423a and a second determination unit 423b.

[0197] The first decision unit 423a modifies the quality information of the training target image according to the training probability distribution using random numbers, and determines the training quality information.

[0198] More specifically, the first decision unit 423a may have the same functions as the quality estimation unit 111. That is, the first decision unit 423a may estimate the position of the target region in the training target image and estimate the quality of the training target image.

[0199] Then, the first decision unit 423a modifies the estimated quality of the training target image according to a training probability distribution using random numbers, and determines the training quality information. The training probability distribution is, for example, a probability distribution in which the highest probability is given to the quality indicated by the quality information of the training target image, and the probability gradually decreases as the degree of difference from that quality increases.

[0200] The second decision unit 423b determines a network policy for the training target image according to the training quality information determined by the first decision unit 423a.

[0201] The information processing device 403 may be physically configured similarly to the information processing device 102 (see Figure 6). However, the storage device 1040 of the information processing device 403 stores program modules for realizing the functions of the information processing device 403. In addition, the network interface 1050 of the information processing device 403 is an interface for connecting the information processing device 403 to a network.

[0202] (Example of operation of the information processing system 400 according to Embodiment 4) The information processing system 400 performs information processing. The information processing according to this embodiment includes learning processing in addition to processing similar to that in Embodiment 1. The learning processing is, for example, processing for training the neural network NN1. For example, when the information processing device 403 receives instructions from the user, it starts the learning processing.

[0203] Figure 16 is a flowchart showing an example of the learning process according to Embodiment 4.

[0204] The learning policy determination unit 423 acquires a learning target image from the learning memory unit 421 (step S401).

[0205] Note that the learning policy determination unit 423 may acquire a learning target image from another information processing device (not shown) via a network or the like, or may acquire a learning target image via a storage medium.

[0206] The learning policy determination unit 423 determines a network policy for the learning target image acquired in step S401 based on the quality information of the learning target image and the learning probability distribution (step S402).

[0207] Specifically, for example, the first determination unit 423a changes the quality information of the learning target image according to the learning probability distribution to determine learning quality information (step S402a).

[0208] more specifically, for example, the first determination unit 423a estimates the quality of the learning target image. Note that the quality of the learning target image may be stored in advance in the learning memory unit 421 in association with the learning target image. In this case, the first determination unit 423a does not have to estimate the quality of the learning target image.

[0209] The first determination unit 423a changes the estimated quality for the learning target image according to the learning probability distribution to determine learning quality information.

[0210] FIG. 17 is a diagram showing an example of the learning probability distribution according to Embodiment 4. The horizontal axis of the figure indicates the quality Q. The vertical axis of the figure indicates the appearance probability P1.

[0211] Q1 in the figure is the quality indicated by the quality information of the learning target image. The learning probability distribution illustrated in the figure is a probability distribution in which the quality Q1 indicated by the quality information of the learning target image appears with the highest probability and gradually appears with a lower probability as the degree of difference from the quality increases. Also, the learning probability distribution shown in the figure is an example of a symmetric probability distribution via the quality Q1 of the learning target image.

[0212] The first decision unit 423a modifies the quality Q1 of the training target image by generating random numbers according to such a training probability distribution. Q2 in the figure is, for example, an example of training quality obtained based on the quality Q1 of the training target image and random numbers generated according to the training probability distribution. In this case, the first decision unit 423a modifies the quality Q1 of the training target image and determines training quality information including training quality Q2.

[0213] Note that the training probability distribution is not limited to the probability distribution exemplified in Figure 17. For example, generally, the lower the quality of the target image, the more difficult it is to authenticate accurately using that image. Therefore, the training probability distribution may be one in which a low quality image (lower than the training target image quality Q1) has a higher probability than a high quality image (higher than the training target image quality Q1). Alternatively, for example, the training probability distribution may be one in which the high quality image has a higher probability than the low quality image. Furthermore, the training policy determination unit 423 may determine multiple network policies for a single training target image.

[0214] Refer to Figure 16 again. The second decision unit 423b determines a network policy for the training target image according to the training quality information determined in step S402a (step S402b).

[0215] In detail, for example, the second decision unit 423b may determine the network policy for the target images for training based on the training quality information determined in step S402a.

[0216] The learning unit 422 inputs the training target image acquired in step S401 into the neural network NN1, which corresponds to the network policy determined in step S402b. As a result, the learning unit 422 trains the neural network NN1 (step S403) and terminates the training process.

[0217] In the learning unit 422, general machine learning techniques should be used for learning. For example, the learning unit 422 (feature extraction unit 422a) extracts features from the training target image using a neural network NN1 according to the network policy (step S403a). Then, the learning unit 422 should estimate the features of the training target image using a neural network NN1 according to the network policy.

[0218] Then, the learning unit 422 (first loss acquisition unit 422b) calculates, for example, the error (first loss) between the feature quantities extracted in step S403a and the correct values ​​of the training target image (step S403b). This allows the learning unit 422 to determine a first loss function for calculating the error (loss) between the estimated feature quantities and the correct values ​​of the training target image. Possible loss functions include, for example, cross-entropy (error), mean squared error, and mean absolute error (MAE).

[0219] The learning unit 422 (first parameter update unit 422c) may calculate the gradient based on the first loss function to determine the first parameter to be used for updating the parameters included in the neural network NN1. Then, the learning unit 422 (first parameter update unit 422c) may update the parameters included in the neural network NN1 using the first parameter (step S403c).

[0220] The learning policy determination unit 423 may acquire multiple training target images in step S401. In this case, the learning policy determination unit 423 and the learning unit 422 may execute steps S402 and S403 for each of the training target images.

[0221] (Effects / Actions) As described above, according to this embodiment, the information processing system 400 further includes a learning unit 422 that performs training on the neural network NN1 by inputting training target images to the neural network NN1.

[0222] This allows us to train the neural network NN1.

[0223] According to this embodiment, the information processing system 400 further includes a learning policy determination unit 423 that determines a network policy for the training target image based on quality information and a learning probability distribution of the training target image. The learning unit 422 learns the neural network NN1 by inputting the training target image to the neural network NN1 according to the determined network policy. The learning probability distribution is a probability distribution in which the probability is highest for the quality indicated by the quality information of the training target image, and gradually decreases as the degree of difference from that quality increases.

[0224] Generally, the quality estimation unit 111 may incorrectly estimate the quality of the target image. According to this embodiment, based on the training target image, it is possible to train the neural network NN1 corresponding to the training quality, which mainly includes the quality of the training target image and the quality of its surroundings. This makes it possible to train the neural network NN1 for image processing (for example, feature extraction using the neural network NN1) that can perform authentication with high accuracy even if the quality of the target image is incorrectly estimated. Therefore, it becomes possible to perform authentication with high accuracy regardless of the quality of the target image.

[0225] According to this embodiment, the learning policy determination unit 423 includes a first determination unit 423a and a second determination unit 423b. The first determination unit 423a modifies the quality information of the target images for training according to a learning probability distribution using random numbers and determines the learning quality information. The second determination unit 423b determines a network policy for the target images for training according to the determined learning quality information.

[0226] This allows the neural network NN1 to be trained based on the target image used for training, primarily by considering the quality of the target image and its surroundings. As a result, as described above, it becomes possible to perform authentication with high accuracy regardless of the quality of the target image.

[0227] <Embodiment 5> In Embodiment 5, another example of the learning method of neural networks NN1 to NN3 will be described. In this embodiment, the learning method of the neural network NN1 according to Embodiment 1 will be described as an example. Also, in this embodiment, for the sake of simplicity, descriptions overlapping with the above-described embodiments will be omitted as appropriate.

[0228] (Configuration example of information processing system 500 according to Embodiment 5) FIG. 18 is a diagram showing a configuration example of an information processing system 500 according to Embodiment 5. The information processing system 500 includes an information processing device 503 in addition to the imaging device 101 and the information processing device 102 similar to those in Embodiment 1.

[0229] The information processing device 503 is a device for performing learning of the neural network NN1. The information processing device 503 includes a learning unit 422 substantially the same as that in Embodiment 4, a learning storage unit 521, an image selection unit 524, and a learning policy determination unit 525.

[0230] The learning storage unit 521 is a storage unit for storing learning data in advance. The learning data according to this embodiment may associate a plurality of learning target images with their respective correct values. The plurality of learning target images may include learning target images with different qualities. The learning data according to this embodiment may further associate the quality of each learning target image related as well.

[0231] The learning unit 422 performs learning of the neural network NN1 by inputting the learning target image selected by the image selection unit 524 (described in detail later) into the neural network NN1.

[0232] The image selection unit 524 selects learning target images such that more learning target images with lower quality are selected. For example, the image selection unit 524 selects learning target images such that more low-resolution learning target images are selected than high-resolution learning target images.

[0233] The learning policy determination unit 525 determines the network policy for the training target images selected by the image selection unit 524 based on the quality information of the training target images selected by the image selection unit 524. The learning policy determination unit 525 processes the policy determination unit 112 to determine the network policy. and A similar process can be used to determine the network policy for the training target images selected by the image selection unit 524.

[0234] The information processing device 503 may be physically configured in the same way as the information processing device 102 (see Figure 6). However, the storage device 1040 of the information processing device 503 stores program modules for realizing the functions of the information processing device 503. Also, the network interface 1050 of the information processing device 503 is an interface for connecting the information processing device 503 to a network.

[0235] (Example of operation of the information processing system 500 according to Embodiment 5) The information processing system 500 performs information processing. The information processing according to this embodiment includes learning processing in addition to processing similar to that in Embodiment 1. The learning processing is, for example, processing for training the neural network NN1, similar to Embodiment 4. For example, when the information processing device 503 receives instructions from the user, it starts the learning processing.

[0236] Figure 19 is a flowchart showing an example of the learning process according to Embodiment 5.

[0237] The image selection unit 524 selects training target images from the training target images stored in the training memory unit 521 in such a way that more training target images of lower quality are selected (step S501).

[0238] In detail, for example, the image selection unit 524 may determine the quality according to a selection probability distribution. Quality is, for example, the resolution of the image.

[0239] Figure 20 shows an example of a selection probability distribution according to Embodiment 5. The horizontal axis of the figure represents quality Q. The vertical axis of the figure represents the probability of occurrence P2. The selection probability distribution is a probability distribution in which lower quality has a higher probability within a predetermined range of quality (in Figure 20, the range of Q3 to Q4). Such a selection probability distribution may be obtained, for example, by multiplying the quality (e.g., resolution) by a predetermined weight.

[0240] The image selection unit 524 then selects a training target image associated with the desired quality from among the training target images stored in the training memory unit 521.

[0241] If the training data does not include quality information, the image selection unit 524 may, for example, estimate the location of the target region in the training image and estimate the quality of the training image, similar to the quality estimation unit 111. In this case, the image selection unit 524 may, for example, select a training image for which the same quality as the quality obtained according to the selection probability distribution has been estimated.

[0242] Refer to Figure 19 again. The image selection unit 524 acquires the training target image selected in step S501 from, for example, the training memory unit 421 (step S502).

[0243] The learning policy determination unit 525 determines the network policy for the learning target images acquired in step S502 based on the quality information of the learning target images acquired in step S502 (step S503).

[0244] In detail, for example, the learning policy determination unit 525 may acquire the quality information obtained in step S501 from the image selection unit 524. The information processing device 503 may also have the same functions as the quality estimation unit 111, and may use these functions to estimate the quality information of the training target image acquired in step S502.

[0245] The learning unit 422 performs training on the neural network NN1, similar to Embodiment 4 (step S403), and then terminates the training process.

[0246] In step S403 of this embodiment, the learning unit 422 inputs the training target image acquired in step S502 to the neural network NN1 according to the network policy determined in step S504. As a result, the learning unit 422 trains the neural network NN1.

[0247] (Effects / Actions) As described above, according to this embodiment, the information processing system 500 further includes an image selection unit 524 that selects training target images in such a way that lower-quality training target images are selected more frequently. The learning unit 422 inputs the selected training target images to the neural network NN1 and performs training on the neural network NN1.

[0248] Generally, the lower the quality of the target image, the more difficult it is to perform accurate authentication using that image. According to this embodiment, since lower-quality training images are selected more frequently, the neural network NN1, which is better suited to lower quality, can undergo more training. This allows for training in image processing (e.g., feature extraction using the neural network NN1) that enables accurate authentication even with low-quality target images. Therefore, accurate authentication becomes possible regardless of the quality of the target image.

[0249] <Embodiment 6> The learning methods described in Embodiments 4 and 5 may be combined. Embodiment 6 describes an example of combining the learning methods described in Embodiments 4 and 5 as yet another example of a learning method for neural networks NN1 to NN3. In this embodiment, the learning method for neural network NN1 according to Embodiment 1 will be described as an example. Also, in this embodiment, in order to simplify the explanation, explanations that overlap with the embodiments described above will be omitted as appropriate.

[0250] (Example of the configuration of the information processing system 600 according to Embodiment 6) Figure 21 shows an example of the configuration of an information processing system 600 according to Embodiment 6. The information processing system 600 includes an imaging device 101 and an information processing device 102 similar to those in Embodiment 1, as well as an information processing device 603.

[0251] The information processing device 603 is a device for training the neural network NN1. The information processing device 603 comprises a learning unit 422 and a learning policy determination unit 423, which are generally the same as in Embodiment 4, and a learning storage unit 521 and an image selection unit 524, which are generally the same as in Embodiment 5.

[0252] In this embodiment, the learning policy determination unit 423 determines a network policy for the learning target image selected by the image selection unit 524 based on the quality information and learning probability distribution of the learning target image.

[0253] The learning policy determination unit 423 may include a first determination unit 423a and a second determination unit 423b, as described in Embodiment 4. That is, the first determination unit 423a may use random numbers to change the quality information of the learning target images selected by the image selection unit 524 according to the learning probability distribution, and determine the learning quality information. The second determination unit 423b may determine the network policy for the learning target images according to the learning quality information determined by the first determination unit 423a.

[0254] The information processing device 603 may be physically configured similarly to the information processing device 102 (see Figure 6). However, the storage device 1040 of the information processing device 603 stores program modules for realizing the functions of the information processing device 603. In addition, the network interface 1050 of the information processing device 603 is an interface for connecting the information processing device 603 to a network.

[0255] (Example of operation of the information processing system 600 according to Embodiment 6) The information processing system 600 performs information processing. The information processing according to this embodiment includes learning processing in addition to processing similar to that in Embodiment 1. The learning processing is, for example, processing for training the neural network NN1, similar to Embodiment 4. For example, when the information processing device 603 receives instructions from the user, it starts the learning processing.

[0256] Figure 22 is a flowchart showing an example of the learning process according to Embodiment 6.

[0257] The image selection unit 524 performs steps S501 and S502, similar to those in Embodiment 5.

[0258] The learning policy determination unit 423 performs a learning policy determination process (step S402) that is generally the same as in Embodiment 4.

[0259] In step S402 of this embodiment, the training target image to be processed is, for example, the training target image acquired in step S502. That is, in step S402 of this embodiment, the training policy determination unit 423 determines a network policy for the training target image acquired in step S502.

[0260] The quality information of the training target image used in step S402 according to this embodiment may be the quality information obtained in step S501. In this case, the training policy determination unit 423 may obtain the quality information of the training target image from the image selection unit 524. The information processing device 603 may also have the same functions as the quality estimation unit 111, and may use these functions to estimate the quality information of the training target image obtained in step S502. In this case, the estimated quality information may be used as the quality of the training target image in the processing in step S402.

[0261] The learning unit 422 performs the process of step S403, which is generally the same as in Embodiment 4, and terminates the learning process.

[0262] In detail, for example, the learning unit 422 may input the training target image acquired in step S502 into the neural network NN1 according to the network policy determined in step S402b. The learning unit 422 then performs training on the neural network NN1.

[0263] (Effects / Actions) As described above, according to this embodiment, the information processing system 600 comprises an image selection unit 524, a learning policy determination unit 423, and a learning unit 422.

[0264] The image selection unit 524 selects training images such that lower-quality training images are selected more frequently.

[0265] The training policy determination unit 423 determines a network policy for the training target image based on the quality information of the training target image and the training probability distribution. The training probability distribution is a probability distribution in which the highest probability is given for the quality indicated by the quality information of the training target image, and the probability gradually decreases as the degree of difference from that quality increases.

[0266] The learning unit 422 trains the neural network NN1 by inputting the selected training target images to the neural network NN1 according to the determined network policy.

[0267] As a result, as described in Embodiment 4, it is possible to train for image processing (e.g., feature extraction using a neural network NN1) that can perform authentication with high accuracy even when the quality of the target image is incorrectly estimated. Furthermore, as described in Embodiment 5, it is possible to train for image processing (e.g., feature extraction using a neural network NN1) that can perform authentication with high accuracy even when the quality of the target image is low. Therefore, it becomes possible to perform authentication with even higher accuracy regardless of the quality of the target image.

[0268] <Embodiment 7> The network policy may be determined using a policy decision model. In this embodiment, an example is described in which the policy decision model is a first policy decision model configured using a neural network. In this embodiment, in order to simplify the explanation, explanations that overlap with the embodiments described above will be omitted as appropriate.

[0269] (Example of the configuration of the information processing system 700 according to Embodiment 7) Figure 23 shows an example configuration of the information processing system 700 according to Embodiment 7. The information processing system 700 includes an imaging device 101 similar to that of Embodiment 1, and an information processing device 702 that replaces the information processing device 102 according to Embodiment 1.

[0270] The information processing device 702 includes, for example, a quality estimation unit 111, an image processing unit 113, and an authentication unit 114, similar in function to those in Embodiment 1, and a policy determination unit 712 that replaces the policy determination unit 112 according to Embodiment 1.

[0271] The policy determination unit 712 determines a network policy based on the quality information estimated by the quality estimation unit 111, similar to the policy determination unit 112 in Embodiment 1. This network policy is applied to the neural network used by the image processing unit 113 for image processing (i.e., image processing of the target image).

[0272] The policy determination unit 712 according to this embodiment determines a network policy according to the quality information of the target image using a trained first policy determination model. The first policy determination model is a machine learning model configured using a neural network.

[0273] The policy determination unit 712 functionally includes, for example, a likelihood acquisition unit 712a and an applicable policy determination unit 712b.

[0274] The likelihood acquisition unit 712a uses a trained first policy decision model to determine the likelihood for each of a predetermined number of network policies according to the quality information of the target image. In other words, the first policy decision model according to this embodiment is a machine learning model that takes the quality information of the target image as input and outputs the likelihood of each network policy.

[0275] In detail, for example, the likelihood acquisition unit 712a inputs the quality information estimated by the quality estimation unit 111 into a pre-trained first policy decision model to obtain the likelihood of each network policy.

[0276] The application policy determination unit 712b determines the network policy to be applied to the image processing of the target image based on the likelihood of each network policy obtained by the likelihood acquisition unit 712a.

[0277] In detail, for example, the application policy determination unit 712b may determine the network policy corresponding to the highest likelihood as the network policy to be applied to the image processing of the target image. Alternatively, for example, the application policy determination unit 712b may probabilistically determine the network policy to be applied to the image processing of the target image according to the likelihood.

[0278] Furthermore, the methods for determining network policies based on the likelihood of each network policy are not limited to these.

[0279] The information processing system 700 according to Embodiment 7 may be physically configured in the same way as the information processing system 100 according to Embodiment 1. However, the storage device 1040 of the information processing device 702 stores program modules for realizing the functions of the information processing device 702. Also, the network interface 1050 of the information processing device 702 is an interface for connecting the information processing device 702 to a network.

[0280] (Example of operation of the information processing system 700 according to Embodiment 7) Figure 24 is a flowchart showing an example of information processing according to Embodiment 7. As shown in the figure, the information processing according to this embodiment includes a step S704 that replaces step S104.

[0281] The policy determination unit 712 determines the network policy based on the quality of the target image estimated in step S103 (step S704).

[0282] Figure 25 is a flowchart showing an example of the network policy determination process (step S704) according to Embodiment 7.

[0283] The likelihood acquisition unit 712a uses the trained first policy decision model and the quality information of the target image estimated in step S103 to determine the likelihood of each of the defined network policies (step S704a).

[0284] The policy application determination unit 712b determines the network policy to be applied to the image processing of the target image based on the likelihood of each network policy obtained in step S704a (step S704b), and returns to information processing.

[0285] (Effects / Actions) As described above, according to this embodiment, the policy determination unit 712 determines a network policy according to the quality information of the target image using a trained first policy determination model. The first policy determination model is configured using a neural network.

[0286] This allows for the determination of a network policy based on the quality of the target image. The processing in the neural network NN1 can then be modified according to this determined network policy. Therefore, authentication can be performed regardless of the quality of the target image while minimizing the increase in processing time.

[0287] <Embodiment 8> Embodiment 8 describes an example of a learning method for the first policy decision model described in Embodiment 7. In this embodiment, in order to simplify the explanation, explanations that overlap with the embodiments described above will be omitted as appropriate.

[0288] (Example of the configuration of the information processing system 800 according to Embodiment 8) Figure 26 shows an example configuration of the information processing system 800 according to Embodiment 8. The information processing system 800 includes an information processing device 803 in addition to the same imaging device 101 and information processing device 102 as in Embodiment 1.

[0289] The information processing device 803 is a device for training the first policy decision model and the neural network NN1. The information processing device 803 includes a learning storage unit 421 and a learning unit 422, similar to those in Embodiment 4, a learning target image acquisition unit 824, and a learning policy decision unit 825.

[0290] The training target image acquisition unit 824 determines the quality of the training target image and acquires the training target image based on the training data stored in the training storage unit 421.

[0291] In detail, for example, the training target image acquisition unit 824 includes a training quality determination unit 824a and an image acquisition unit 824b.

[0292] The learning quality determination unit 824a randomly determines the quality of the training target image according to a predetermined learning probability distribution (see, for example, Embodiment 4).

[0293] The image acquisition unit 824b acquires training target images according to the quality determined by the training quality determination unit 824a. For example, the image acquisition unit 824b may acquire images of randomly determined quality from the training storage unit 421. Alternatively, for example, the image acquisition unit 824b may acquire an image from the training storage unit 421 and acquire training target images by degrading the quality of the acquired image according to a randomly determined quality.

[0294] The learning policy determination unit 825 trains the first policy determination model. For example, the learning policy determination unit 825 calculates the likelihood of each of several network policies. Based on the likelihood of each network policy, the learning policy determination unit 825 determines the network policy to be applied to the image processing of the training target image. In this case, the learning policy determination unit 825 may determine the network policy in the same way as the applied policy determination unit 712b. Based on the first loss, the learning policy determination unit 825 updates the parameters of the first policy determination model.

[0295] The information processing device 803 may be physically configured similarly to the information processing device 102, for example (see Figure 6). However, the storage device 1040 of the information processing device 803 stores program modules for realizing the functions of the information processing device 803. Also, the network interface 1050 of the information processing device 803 is an interface for connecting the information processing device 803 to a network.

[0296] (Example of operation of the information processing system 800 according to Embodiment 8) The information processing system 800 performs information processing. The information processing according to this embodiment includes, for example, a learning process. The learning process is, for example, a process for training a first policy decision model and a neural network NN1. For example, when the information processing device 803 receives instructions from the user, it starts the learning process.

[0297] Figure 27 is a flowchart showing an example of the learning process according to Embodiment 8.

[0298] The training target image acquisition unit 824 acquires training target images of randomly determined quality (step S801).

[0299] In detail, for example, the training quality determination unit 824a randomly determines the quality of the training target image according to a predetermined training probability distribution (step S801a).

[0300] The image acquisition unit 824b acquires a training target image of the quality determined in step S801a, for example, using training data stored in the training memory unit 421 (step S801b).

[0301] The learning policy determination unit 825 uses the first policy determination model to determine the network policy for the training target image acquired in step S801b (step S802).

[0302] The learning unit 422, similar to Embodiment 4, trains the neural network NN1 according to the network policy determined in step S802 (step S403).

[0303] The learning policy determination unit 825 updates the parameters (second parameters) of the first policy determination model (step S804) using the first loss obtained in step S403b, and then terminates the learning process.

[0304] Here, we have described an example of simultaneously training the first policy decision model and the neural network NN1, but the training of the first policy decision model and the neural network NN1 may be performed separately. For example, if you are training only one of the first policy decision model or the neural network NN1, it is a good idea to keep the parameters of the other component fixed.

[0305] (Effects / Actions) As described above, according to this embodiment, the information processing device 803 includes a learning policy determination unit 825 that performs learning of the first policy determination model.

[0306] This allows for the generation of a pre-trained first policy decision model, which can then be used to determine a network policy based on the quality of the target image. The processing in the neural network NN1 can then be modified according to the determined network policy. Therefore, authentication can be performed regardless of the quality of the target image while minimizing the increase in processing time.

[0307] <Embodiment 9> Embodiment 9 describes a detailed example of the learning method for the first policy decision model described in Embodiment 8. In this embodiment, an example is described in which the first policy decision model is learned separately from the learning of the neural network NN1. Also in this embodiment, in order to simplify the explanation, explanations that overlap with the embodiments described above are omitted as appropriate.

[0308] (Example of the configuration of the information processing system according to Embodiment 9) In the information processing system according to this embodiment, the information processing device according to this embodiment includes a learning policy determination unit 925 that replaces the learning policy determination unit 825.

[0309] Figure 28 shows an example of the configuration of the learning policy determination unit 925 according to Embodiment 9. Functionally, the learning policy determination unit 925 includes, for example, a first learning likelihood acquisition unit 925a and a second parameter update unit 925b.

[0310] The first training likelihood acquisition unit 925a inputs the quality of randomly determined training target images into the first policy decision model and calculates the likelihood (vector) for each of the multiple network policies.

[0311] The second parameter update unit 925b updates the parameters (second parameters) of the first policy decision model based on the first loss and likelihood for each of the multiple network policies.

[0312] In detail, for example, the second parameter update unit 925b includes a second loss acquisition unit 925b1 and a second parameter determination unit 925b2.

[0313] The second loss acquisition unit 925b1 calculates the second loss (overall loss), which is the sum of the first losses weighted by likelihood for each of the multiple network policies.

[0314] The second parameter determination unit 925b2 updates the (second parameter) of the first policy determination model based on the second loss.

[0315] The information processing device according to this embodiment may be physically configured similarly to, for example, the information processing device 802 (see Figure 6). However, the storage device 1040 of the information processing device according to this embodiment stores program modules for realizing the functions of the information processing device according to this embodiment. Furthermore, the network interface 1050 of the information processing device according to this embodiment is an interface for connecting the information processing device according to this embodiment to a network.

[0316] (Example of operation of the information processing system 900 according to Embodiment 9) The information processing system 900 performs information processing. The information processing according to this embodiment includes, for example, a learning process. The learning process according to this embodiment includes a first policy decision model learning process for learning a first policy decision model.

[0317] For training the neural network NN1, for example, the training processes of Embodiments 4 and 5 may be applied. In this embodiment, an example is described in which the training of the neural network NN1 is performed separately from the training of the first policy decision model, but the training of the first policy decision model and the neural network NN1 may be performed simultaneously.

[0318] Figure 29 is a flowchart showing an example of the first policy decision model learning process according to Embodiment 9. For example, when the information processing device receives instructions from the user, it starts the first policy decision model learning process.

[0319] For example, first, the same process as in Embodiment 8, step S801, is executed.

[0320] The first learning likelihood acquisition unit 925a calculates the likelihood (vector) for each of the multiple network policies (step S901).

[0321] For example, the first training likelihood acquisition unit 925a inputs the quality of the training target image, which was randomly determined in step S801 (see step S801a in Figure 27), into the first policy decision model to obtain the likelihood (vector) of each network policy.

[0322] The feature extraction unit 422a extracts features from the training target image using the neural network NN1 corresponding to each of the multiple network policies (step S902).

[0323] The first loss acquisition unit 422b calculates the error (first loss) between the feature quantities extracted in step S902 and the correct values ​​of the training target images for each of the multiple network policies (step S903).

[0324] The second parameter update unit 925b updates the parameters (second parameters) of the first policy decision model based on the likelihood and first loss obtained in steps S901 and S902 (step S904).

[0325] In detail, for example, the second loss acquisition unit 925b1 calculates the second loss (overall loss), which is the sum of the first losses weighted by likelihood for each of the multiple network policies (step S904a). Based on this, the second loss acquisition unit 925b1 can then determine the second loss function for calculating the second loss.

[0326] The second parameter determination unit 925b2 updates the (second parameter) of the first policy decision model based on the second loss obtained in step S904a (step S904b), and terminates the first policy decision model learning process.

[0327] For example, the second parameter determination unit 925b2 may calculate the gradient based on the second loss function to determine the second parameter to be used to update the parameters included in the first policy decision model. Then, the second parameter determination unit 925b2 may update the parameters included in the first policy decision model using the determined second parameter. The first policy decision model learning process may be executed repeatedly until predetermined conditions are met.

[0328] (Effects / Actions) As described above, according to this embodiment, the information processing device comprises a learning policy determination unit 925 and a first loss acquisition unit 422b. The learning policy determination unit 925 learns a first policy determination model for determining a network policy corresponding to the quality information of a target image from among predetermined network policies. The first loss acquisition unit 422b calculates a first loss when processing a target image for training using a neural network corresponding to each of the multiple network policies.

[0329] The first policy decision model consists of a neural network.

[0330] The learning policy determination unit 925 includes a first learning likelihood acquisition unit 925a and a second parameter update unit 925b. The first learning likelihood acquisition unit 925a inputs quality information of the target images for training into the first policy determination model and calculates the likelihood of each network policy. The second parameter update unit 925b updates the parameters of the first policy determination model based on the first loss and likelihood of each network policy.

[0331] The second parameter update unit 925b includes a second loss acquisition unit 925b1 and a second parameter determination unit 925b2. The second loss acquisition unit 925b1 calculates a second loss for each network policy, which is the sum of the first losses weighted by likelihood. The second parameter determination unit 925b2 updates the parameters of the first policy determination model based on the second loss.

[0332] This allows for the generation of a pre-trained first policy decision model, which can then be used to determine a network policy based on the quality of the target image. The processing in the neural network NN1 can then be modified according to the determined network policy. Therefore, authentication can be performed regardless of the quality of the target image while minimizing the increase in processing time.

[0333] <Embodiment 10> Embodiment 7 describes an example in which the network policy is determined using a first policy decision model configured with a neural network. A neural network is not required for the policy decision model used to determine the network policy. In this embodiment, the policy decision model used to determine the network policy may consist of multiple probability distributions that include learnable parameters. The multiple probability distributions may correspond to each of a predetermined set of network policies.

[0334] In this embodiment, a policy decision model composed of multiple probability distributions is referred to as the second policy decision model. Furthermore, the probability distributions constituting the second policy decision model are referred to as policy probability distributions.

[0335] Furthermore, in this embodiment, in order to simplify the explanation, explanations that overlap with the embodiments described above will be omitted as appropriate.

[0336] (Example of the configuration of the information processing system 1100 according to Embodiment 10) Figure 30 shows an example configuration of the information processing system 1100 according to Embodiment 10. The information processing system 1100 includes an imaging device 101 similar to that of Embodiment 1, and an information processing device 1102 that replaces the information processing device 102 according to Embodiment 1.

[0337] The information processing device 1102 includes, for example, a quality estimation unit 111, an image processing unit 113, and an authentication unit 114, similar in function to those in Embodiment 1, and a policy determination unit 1112 that replaces the policy determination unit 112 in Embodiment 1.

[0338] The policy determination unit 1112 determines a network policy based on the quality information estimated by the quality estimation unit 111, similar to the policy determination unit 112 in Embodiment 1. This network policy is applied to the neural network used by the image processing unit 113 for image processing (i.e., image processing of the target image).

[0339] The policy determination unit 1112 according to this embodiment determines a network policy according to the quality information of the target image using a trained second policy determination model. The second policy determination model is, for example, a model for determining a network policy according to the quality information of the target image from among predetermined network policies. The second policy determination model consists of multiple policy probability distributions. Each of the multiple policy probability distributions corresponds to each of the multiple predetermined neural networks.

[0340] The policy determination unit 1112 functionally includes, for example, a likelihood acquisition unit 1112a and an application policy determination unit 712b similar to that in Embodiment 7.

[0341] The likelihood acquisition unit 1112a uses a trained second policy decision model to determine the likelihood for each of a predetermined number of network policies according to the quality information of the target image.

[0342] The second policy decision model according to this embodiment consists of multiple policy probability distributions associated with each of the multiple network policies. The likelihood acquisition unit 1112a according to this embodiment calculates the likelihood of each of the multiple policy probability distributions.

[0343] The information processing system 1100 according to Embodiment 10 may be physically configured in the same way as, for example, the information processing system 100 according to Embodiment 1. However, the storage device 1040 of the information processing device 1102 stores program modules for realizing the functions of the information processing device 1102. Also, the network interface 1050 of the information processing device 1102 is an interface for connecting the information processing device 1102 to a network.

[0344] (Example of operation of the information processing system 1100 according to Embodiment 10) Figure 31 is a flowchart showing an example of information processing according to Embodiment 10. As shown in the figure, the information processing according to this embodiment includes a step S1104 that replaces step S104.

[0345] The policy determination unit 1112 determines the network policy based on the quality of the target image estimated in step S103 (step S1104).

[0346] Figure 32 is a flowchart showing an example of the network policy determination process (step S1104) according to Embodiment 10.

[0347] The likelihood acquisition unit 1112a uses the trained second policy decision model and the quality information of the target image estimated in step S103 to determine the likelihood of each of the predetermined network policies (step S1104a).

[0348] In detail, for example, the likelihood acquisition unit 1112a uses the quality information of the target image estimated in step S103 to determine the likelihood of each of the multiple policy probability distributions that constitute the second policy decision model.

[0349] The policy application unit 712b, similar to Embodiment 7, determines the network policy to be applied to the image processing of the target image based on the likelihood of each network policy obtained in step S1104a (step S704b), and returns to information processing.

[0350] In other words, for example, the application policy determination unit 712b may determine the network policy corresponding to the policy probability distribution with the highest likelihood for the quality of the target image as the network policy to be applied to the image processing of the target image.

[0351] Alternatively, for example, the application policy determination unit 712b may probabilistically determine the network policy to be applied to the image processing of the target image based on the likelihood of each policy probability distribution for the quality of the target image. Specifically, for example, suppose the likelihood when the quality of the target image is x is 0.2 for policy probability distribution A, 0.5 for policy probability distribution B, and 0.3 for policy probability distribution C. In this case, the application policy determination unit 712b may, for example, select network policies A, B, and C corresponding to policy probability distributions A, B, and C, respectively, with probabilities of 20%, 50%, and 30%, respectively.

[0352] (Effects / Actions) As described above, according to this embodiment, the policy determination unit 1112 determines a network policy according to the quality information of the target image using a trained second policy determination model. The second policy determination model consists of multiple policy probability distributions associated with each of the multiple network policies.

[0353] This allows for the determination of a network policy based on the quality of the target image. The processing in the neural network NN1 can then be modified according to this determined network policy. Therefore, authentication can be performed regardless of the quality of the target image while minimizing the increase in processing time.

[0354] <Embodiment 11> Embodiment 11 describes an example of a learning method for the second policy decision model described in Embodiment 10. In this embodiment, in order to simplify the explanation, explanations that overlap with the embodiments described above will be omitted as appropriate.

[0355] (Example of the configuration of the information processing system 1200 according to Embodiment 11) Figure 33 shows an example configuration of the information processing system 1200 according to Embodiment 11. The information processing system 1200 includes an imaging device 101 and an information processing device 102 similar to those in Embodiment 1, as well as an information processing device 1203.

[0356] The information processing device 1203 is a device for training the second policy decision model and the neural network NN1. The information processing device 1203 includes a learning storage unit 421 and a learning unit 422 similar to those in Embodiment 4, a learning target image acquisition unit 824 similar to those in Embodiment 8, and a learning policy decision unit 1225.

[0357] The learning policy determination unit 1225 performs training on the second policy determination model. For example, the learning policy determination unit 1225 calculates the likelihood of each of the multiple network policies, that is, the likelihood of each of the multiple policy probability distributions that constitute the second policy determination model.

[0358] The learning policy determination unit 1225 determines the network policy to be applied to the image processing of the training target image based on the likelihood of each policy probability distribution. In this case, for example, the learning policy determination unit 1125 may determine the network policy in the same way as the applied policy determination unit 712b. The learning policy determination unit 1225 updates the parameters (third parameters) of the second policy determination model (third parameters) based, for example, the first loss.

[0359] Figure 34 shows an example of the functional configuration of the learning policy determination unit 1225 according to Embodiment 11. The learning policy determination unit 1225 includes, for example, a second learning likelihood acquisition unit 1225a and a third parameter update unit 1225b.

[0360] The second training likelihood acquisition unit 1225a calculates the likelihood (vector) for each of the multiple network policies based on the quality of the training target images, which are randomly determined, and the second policy decision model.

[0361] For example, the second training likelihood acquisition unit 1225a uses the quality of randomly determined training target images to calculate the likelihood of each of the multiple policy probability distributions that constitute the second policy decision model.

[0362] The third parameter update unit 1225b updates the parameters of the second policy decision model based on the first loss and the modified loss.

[0363] The corrected loss is a loss that adjusts for the difference in parameters of policy probability distributions corresponding to different network policies so that it is large. For example, if the mean of policy probability distributions A and B is μ A and μ B In this case, the corrected loss is μ A and μ B The reciprocal of the absolute difference (1 / |μ) A -μ B |) Note that the adjusted loss is not limited to this.

[0364] The parameters of the second policy decision model are one or more parameters included in each of the multiple policy probability distributions. Examples of parameters include the variance and mean when the policy probability distribution is normally distributed. However, the parameters are not limited to these, and the policy probability distribution is not limited to a normally distributed distribution.

[0365] In detail, for example, the third parameter update unit 1225b includes a third loss acquisition unit 1225b1 and a third parameter determination unit 1225b2.

[0366] The third loss acquisition unit 1225b1 calculates the third loss by adding the first loss and the adjusted loss. However, the third loss based on the first loss and the adjusted loss is not limited to this.

[0367] The third parameter determination unit 1225b2 updates the parameters of each policy probability distribution based on the third loss.

[0368] The information processing device 1203 according to this embodiment may be physically configured similarly to, for example, the information processing device 802 (see Figure 6). However, the storage device 1040 of the information processing device 1203 stores program modules for realizing the functions of the information processing device 1203. In addition, the network interface 1050 of the information processing device 1203 is an interface for connecting the information processing device 1203 to a network.

[0369] (Example of operation of the information processing system 1200 according to Embodiment 11) The information processing system 1200 performs information processing. The information processing according to this embodiment includes learning processing. The learning processing is, for example, the process for training the first policy decision model and the neural network NN1.

[0370] In this embodiment, an example is described in which a second policy decision model is trained separately from the training of the neural network NN1. For example, the training processes of Embodiments 4 and 5 may be applied to the training of the neural network NN1. Alternatively, the second policy decision model may be applied instead of the first policy decision model described in Embodiment 8, and the training of the second policy decision model and the neural network NN1 may be performed simultaneously.

[0371] Figure 35 is a flowchart showing an example of the second policy decision model learning process according to Embodiment 11. For example, when the information processing device receives instructions from the user, it starts the second policy decision model learning process.

[0372] For example, first, the same process as in Embodiment 8, step S801, is executed.

[0373] The second learning likelihood acquisition unit 1225a calculates the likelihood (vector) for each of the multiple network policies (step S1201).

[0374] For example, the second learning likelihood acquisition unit 1225a calculates the likelihood of each policy probability distribution corresponding to the quality of the training target image randomly determined in step S801 (see step S801a in Figure 27). The second learning likelihood acquisition unit 1225a may determine the network policy based on the likelihood in the same manner as the applied policy determination unit 712b.

[0375] The feature extraction unit 422a and the first loss acquisition unit 422b each perform the same steps S902 and S903 as in Embodiment 9. If a network policy is determined in step S901, the neural network is used to obtain the feature quantities and the first loss according to the network policy.

[0376] The third parameter update unit 1225b updates the parameters (third parameters) of the second policy decision model based on the first loss obtained in step S902 (step S1204).

[0377] In detail, for example, the third loss acquisition unit 1225b1 calculates the third loss by adding the first loss and the corrected loss obtained in step S902 (step S1204a). This allows the third loss acquisition unit 1225b1 to determine the third loss function for calculating the third loss.

[0378] The third parameter determination unit 1225b2 updates the parameters (third parameters) of each policy probability distribution based on the third loss obtained in step S1204a (step S1204b), and terminates the second policy decision model learning process.

[0379] For example, the third parameter determination unit 1225b2 may calculate the gradient based on the third loss function to determine the third parameter to be used for updating the parameters included in the second policy decision model (each policy probability distribution). Then, the third parameter determination unit 1225b2 may update the parameters included in the second policy decision model (each policy probability distribution) using the determined third parameter. The second policy decision model learning process may be executed repeatedly until predetermined conditions are met.

[0380] (Effects / Actions) As described above, according to this embodiment, the information processing device 1203 comprises a learning policy determination unit 1225 and a first loss acquisition unit 422b. The learning policy determination unit 1225 learns a second policy determination model for determining a network policy corresponding to the quality information of a target image from among predetermined network policies. The first loss acquisition unit 422b calculates a first loss when processing a target image for training using a neural network corresponding to the network policy determined using the second policy determination model.

[0381] The second policy decision model consists of multiple policy probability distributions, each associated with a different network policy. Each of these policy probability distributions contains one or more parameters.

[0382] The learning policy determination unit 1225 includes a second learning likelihood acquisition unit 1225a and a third parameter update unit 1225b. The second learning likelihood acquisition unit 1225a uses quality information of the target image for training and each of the multiple policy probability distributions to determine the likelihood of each policy probability distribution. The third parameter update unit 1225b updates the parameters of each policy probability distribution based on a first loss and a modified loss that corrects the difference between the parameters of the policy probability distributions corresponding to different network policies to be large.

[0383] The third parameter update unit 1225b includes a third loss acquisition unit 1225b1 and a third parameter determination unit 1225b2. The third loss acquisition unit 1225b1 calculates the third loss by adding the first loss and the modified loss. The third parameter determination unit 1225b2 updates the parameters of each policy probability distribution based on the third loss.

[0384] This allows for the generation of a trained second policy decision model, which can then be used to determine a network policy based on the quality of the target image. The processing in the neural network NN1 can then be modified according to the determined network policy. Therefore, authentication can be performed regardless of the quality of the target image while minimizing the increase in processing time.

[0385] Furthermore, since high-quality images are generally easier to process, updating the parameters of each policy probability distribution based solely on the first loss may lead to a training probability distribution that samples high-quality training images at a high frequency. In this embodiment, a corrected loss is added to update the parameters of each policy probability distribution. This prevents the training probability distribution from becoming one that samples high-quality training images at a high frequency. Therefore, it becomes possible to train the second policy decision model using training images of various qualities selected with appropriate probabilities.

[0386] <Embodiment 12> Embodiment 12 describes an example of a third policy decision model that determines the network policy using the target image itself as input instead of the quality of the target image. The third policy decision model is configured, for example, using a neural network. In this embodiment, in order to simplify the explanation, explanations that overlap with the embodiments described above will be omitted as appropriate.

[0387] (Example of the configuration of the information processing system 1300 according to Embodiment 12) Figure 36 shows an example configuration of the information processing system 1300 according to Embodiment 12. The information processing system 1300 includes an imaging device 101 similar to that of Embodiment 1, and an information processing device 1302 that replaces the information processing device 102 according to Embodiment 1.

[0388] The information processing device 1302 includes, for example, an image processing unit 113 and an authentication unit 114 similar in function to those in Embodiment 1, and a policy determination unit 1312 that replaces the policy determination unit 112 according to Embodiment 1.

[0389] The policy determination unit 1312 determines a network policy based on the target image. This network policy is applied to the neural network used by the image processing unit 113 for image processing (i.e., image processing of the target image).

[0390] The policy determination unit 1312 according to this embodiment determines a network policy corresponding to the target image using a trained third policy determination model. The third policy determination model is, for example, a machine learning model for determining a network policy corresponding to the target image from among a predetermined number of network policies. The third policy determination model is configured using a neural network.

[0391] The policy determination unit 1312 functionally includes, for example, a likelihood acquisition unit 1312a and an application policy determination unit 712b similar to that in Embodiment 7.

[0392] The likelihood acquisition unit 1312a uses a trained third policy decision model to determine the likelihood for each of a predetermined number of network policies corresponding to the target image. In other words, the third policy decision model according to this embodiment is a machine learning model that takes a target image as input and outputs the likelihood of each network policy.

[0393] In detail, for example, the likelihood acquisition unit 1312a inputs the target image into a pre-trained third policy decision model to obtain the likelihood of each network policy.

[0394] The information processing system 1300 according to Embodiment 12 may be physically configured in the same way as, for example, the information processing system 100 according to Embodiment 1. However, the storage device 1040 of the information processing device 1302 stores program modules for realizing the functions of the information processing device 1302. Also, the network interface 1050 of the information processing device 1302 is an interface for connecting the information processing device 1302 to a network.

[0395] (Example of operation of the information processing system 1300 according to Embodiment 12) Figure 37 is a flowchart showing an example of information processing according to Embodiment 12. As shown in the figure, the information processing according to this embodiment includes step S1304 instead of step S104. Furthermore, the information processing according to this embodiment does not need to include step S103.

[0396] The policy determination unit 1312 determines a network policy based on the target image (step S1304).

[0397] Figure 38 is a flowchart showing an example of the network policy determination process (step S1304) according to Embodiment 12.

[0398] The likelihood acquisition unit 1312a uses the target image and the trained third policy decision model to determine the likelihood of each of the predetermined network policies (step S1304a).

[0399] The policy application unit 712b, similar to Embodiment 7, determines the network policy to be applied to the image processing of the target image based on the likelihood of each network policy obtained in step S1304a (step S704b), and returns to information processing.

[0400] (Effects / Actions) As described above, according to this embodiment, the policy determination unit 1312 determines a network policy corresponding to the target image using a trained third policy determination model. The third policy determination model is configured using a neural network.

[0401] This allows for the determination of a network policy based on the target image. The processing in the neural network NN1 can then be modified according to this determined network policy. Therefore, authentication can be performed regardless of the quality of the target image while minimizing the increase in processing time.

[0402] <Embodiment 13> Embodiment 13 describes an example of a training method for the third policy decision model described in Embodiment 12. In training the third policy decision model, a trained policy decision model (such as the first policy decision model or the second policy decision model) is used. In this embodiment, an example using the first policy decision model for training the third policy decision model is provided.

[0403] Furthermore, in this embodiment, in order to simplify the explanation, explanations that overlap with the embodiments described above will be omitted as appropriate.

[0404] (Example of the configuration of the information processing system 1400 according to Embodiment 13) Figure 39 shows an example of the configuration of the information processing system 1400 according to Embodiment 13. The information processing system 1400 includes an information processing device 1403 in addition to the same imaging device 101 and information processing device 102 as in Embodiment 1.

[0405] The information processing device 1403 is a device for training the third policy decision model and the neural network NN1. The information processing device 1403 includes a learning storage unit 421 and a learning unit 422 similar to those in Embodiment 4, a learning target image acquisition unit 824 similar to those in Embodiment 8, and a learning policy decision unit 1425.

[0406] The learning policy determination unit 1425 uses the trained first policy determination model to train the third policy determination model. For example, the learning policy determination unit 1425 obtains the likelihood (ground truth likelihood) based on the first policy determination model and the likelihood (learning likelihood) based on the third policy determination model for each of the multiple network policies. Then, the learning policy determination unit 1425 trains the third policy determination model so that the learning likelihood approaches the ground truth likelihood.

[0407] Figure 40 shows an example of the functional configuration of the learning policy determination unit 1425 according to Embodiment 13. The learning policy determination unit 1425 includes, for example, a correct likelihood acquisition unit 1425a, a third learning likelihood acquisition unit 1425b, and a fourth parameter update unit 1425c.

[0408] The ground truth likelihood acquisition unit 1425a inputs the quality of randomly selected training target images into a pre-trained first policy decision model and obtains the likelihood (vector) of each of the multiple network policies as the ground truth likelihood.

[0409] The third training likelihood acquisition unit 1425b inputs training target images, randomly determined according to quality, into the third policy decision model and obtains the likelihood (vector) of each network policy as the likelihood during training.

[0410] The fourth parameter update unit 1425c updates the parameters (fourth parameters) of the third policy decision model based on the fourth loss so that the likelihood during training approaches the true likelihood. The fourth loss is, for example, the error between the true likelihood obtained for multiple network policies and the likelihood during training. For the fourth loss, cross-entropy (error), mean squared error, mean absolute error (MAE), etc., may be used.

[0411] In detail, for example, the fourth parameter update unit 1425c includes a fourth loss acquisition unit 1425c1 and a fourth parameter determination unit 1425c2.

[0412] The fourth loss acquisition unit 1425c1 calculates the fourth loss based on the correct likelihood obtained for multiple network policies and the likelihood during training.

[0413] The fourth parameter determination unit 1425c2 updates the parameters of the third policy determination model based on the fourth loss.

[0414] The information processing device 1403 according to this embodiment may be physically configured similarly to, for example, the information processing device 802 (see Figure 6). However, the storage device 1040 of the information processing device 1403 stores program modules for realizing the functions of the information processing device 1403. In addition, the network interface 1050 of the information processing device 1403 is an interface for connecting the information processing device 1403 to a network.

[0415] (Example of operation of the information processing system 1400 according to Embodiment 13) The information processing system 1400 performs information processing. The information processing according to this embodiment includes learning processing. The learning processing is, for example, the process for training the third policy decision model and the neural network NN1.

[0416] In this embodiment, an example is described in which the third policy decision model is trained separately from the training of the neural network NN1. For example, the training processes of Embodiments 4 and 5 may be applied to the training of the neural network NN1. Alternatively, the third policy decision model may be applied instead of the first policy decision model described in Embodiment 8, and the training of the third policy decision model and the neural network NN1 may be performed simultaneously.

[0417] Figure 41 is a flowchart showing an example of the third policy decision model learning process according to Embodiment 13. For example, when the information processing device receives instructions from the user, it starts the third policy decision model learning process.

[0418] For example, first, the same process as in Embodiment 8, step S801, is executed.

[0419] The ground truth likelihood acquisition unit 1425a inputs the quality of the training target image, which was randomly determined in step S801a, into the trained first policy decision model to obtain the ground truth likelihood for each of the multiple network policies (step S1401).

[0420] For example, the ground truth likelihood acquisition unit 1425a pre-stores a trained first policy decision model. The ground truth likelihood acquisition unit 1425a inputs the quality of randomly selected training target images into the pre-stored first policy decision model to obtain the ground truth likelihood (vector) for each of the multiple network policies.

[0421] The third training likelihood acquisition unit 1425b inputs the training target images acquired in step S801b into the third policy decision model to obtain the likelihood of each network policy during training (step S1402).

[0422] For example, the third training likelihood acquisition unit 1425b inputs the training target image acquired in step S801b into the third policy decision model. This allows the third training likelihood acquisition unit 1425b to calculate the training likelihood for each of the multiple network policies.

[0423] The fourth parameter update unit 1425c updates the parameters (fourth parameters) of the third policy decision model (step S1403) so that the correct likelihood obtained in steps S1401 and S1402 approaches the likelihood during training.

[0424] In detail, for example, the fourth loss acquisition unit 1425c1 calculates the fourth loss based on the correct likelihood obtained for multiple network policies and the likelihood during training (step S1403a). This allows the fourth loss acquisition unit 1425c1 to determine the fourth loss function for calculating the fourth loss.

[0425] The fourth parameter determination unit 1425c2 updates the parameters of the third policy decision model based on the fourth loss obtained in step S1403a (step S1403b), and terminates the third policy decision model learning process.

[0426] For example, the fourth parameter determination unit 1425c2 may calculate the gradient based on the fourth loss function to determine the fourth parameter to be used in updating the third policy decision model. Then, the fourth parameter determination unit 1425c2 updates the parameters included in the third policy decision model using the determined fourth parameter. The third policy decision model learning process may be executed repeatedly until predetermined conditions are met.

[0427] (Effects / Actions) As described above, according to this embodiment, the information processing device 1403 further comprises a learning policy determination unit 1425. The learning policy determination unit 1425 learns a third policy determination model for determining a network policy corresponding to a target image from among a predetermined number of network policies.

[0428] The learning policy determination unit 1425 includes a correct likelihood acquisition unit 1425a, a third learning likelihood acquisition unit 1425b, and a fourth parameter update unit 1425c.

[0429] The ground truth likelihood acquisition unit 1425a inputs the quality information of the target image for training into a pre-trained first policy decision model for determining a network policy from a predetermined set of network policies according to the quality information of the target image, and obtains the likelihood of each network policy as the ground truth likelihood. The third training likelihood acquisition unit 1425b inputs the target image for training into the third policy decision model and obtains the likelihood of each network policy as the likelihood during training. The fourth parameter update unit 1425c updates the parameters of the third policy decision model based on the fourth loss so that the likelihood during training approaches the ground truth likelihood.

[0430] The fourth loss is the loss based on the error between the likelihood during training and the true likelihood.

[0431] This allows for the generation of a trained third-policy decision model, which can then be used to determine a network policy corresponding to the target image. The processing in the neural network NN1 can then be modified according to the determined network policy. Therefore, authentication can be performed regardless of the quality of the target image while minimizing the increase in processing time.

[0432] The embodiments and modifications of the present invention have been described above with reference to the drawings, but these are merely examples of the present invention, and various other configurations can also be adopted.

[0433] Furthermore, while the flowcharts used in the above description show multiple steps (processes) in sequence, the execution order of the steps performed in each embodiment is not limited to the order in which they are described. In each embodiment, the order of the illustrated steps can be changed to the extent that it does not impede the content. Also, the above embodiments and modifications can be combined to the extent that their content is not contradictory.

[0434] Some or all of the above embodiments may also be described as follows, but are not limited to the following:

[0435] 1. A quality estimation means for estimating the quality information of the target image, The system comprises an image processing means that processes the target image using a neural network that corresponds to a network policy determined based on the quality information of the target image. Information processing system. 2. The image processing means extracts features from the target image using the neural network according to the network policy. The information processing system described in 1. 3. The aforementioned target image is an eye image that includes the iris region in which the iris is reflected. The aforementioned quality information includes the iris diameter. The information processing system described in 1. or 2. 4. The quality information includes the resolution of the target image. An information processing system described in any one of the following three items. 5. Further comprising a policy determination means for determining the network policy based on the quality information. An information processing system described in any one of items 1 through 4. 6. The policy determination means determines the network policy such that the processing time using the neural network is within a predetermined range, regardless of the quality information of the target image. The information processing system described in 5. 7. The image processing means is: A first processing means that converts the iris region included in the target image into a standard image of a predetermined shape, The system includes a second processing means that processes the standard image using the neural network according to the network policy, The first processing means converts the iris region into a standard image of a size corresponding to the quality information. An information processing system described in any one of items 1 through 6. 8. The neural network is composed of a pre-stage group including an input layer and a post-stage group including a final layer. The aforementioned network policy is applied to image processing using the aforementioned group. An information processing system described in any one of items 1 through 7. 9. The neural network includes multiple convolutional layers, The policy determination means determines the network policy such that the lower the resolution of the target image, the larger the number of convolutional layers included in the preceding group. The information processing system described in 8. 10. The neural network includes multiple convolutional layers, The policy determination means determines the network policy such that the lower the resolution of the target image, the larger the number of channels in the calculations performed in the convolutional layer included in the preceding group. The information processing system described in 8. 11. The system further comprises a learning means for training the neural network by inputting target images for training into the neural network. An information processing system described in any one of items 1 through 10. 12. The system further comprises a training policy determination means for determining a network policy for the training target image based on the quality information and training probability distribution of the training target image, The learning means performs training on the neural network by inputting the target images for training into the neural network according to the determined network policy. The aforementioned training probability distribution is one in which the highest probability is given for the quality indicated by the quality information of the training target image, and the probability gradually decreases as the degree of difference from that quality increases. The information processing system described in 11. 13. The learning policy determination means is: A first determination means that modifies the quality information of the target image for training according to the training probability distribution by using the aforementioned random numbers, and determines the training quality information, Includes a second determination means for determining a network policy for the target images for training based on the determined training quality information. The information processing system described in 12. 14. The system further provides image selection means for selecting training images so that lower-quality training images are selected more frequently. The learning means performs training on the neural network by inputting the selected training target images into the neural network. An information processing system described in any one of items 11 through 13. 15. A learning policy determination means for training a first policy determination model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, The system further comprises a first loss acquisition means for determining the first loss when processing the training target image using a neural network corresponding to each of the aforementioned multiple network policies, The aforementioned first policy decision model consists of a neural network, The aforementioned learning policy determination means is: A first learning likelihood acquisition means inputs the quality information of the target images for training into the first policy decision model and calculates the likelihood of each network policy, The system includes a second parameter updating means for updating the parameters of the first policy decision model based on the first loss and likelihood of each of the aforementioned network policies, The second parameter update means is: A second loss acquisition means for obtaining a second loss which is the sum of the first losses weighted by likelihood for each of the aforementioned network policies, Includes a second parameter update means for updating the parameters of the first policy decision model based on the second loss. The information processing system described in 11. 16. A learning policy determination means for training a second policy determination model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, The system further comprises a first loss acquisition means for calculating the first loss when processing the training target image using a neural network corresponding to the network policy determined using the second policy determination model, The second policy decision model consists of a plurality of policy probability distributions associated with each of the plurality of network policies, Each of the aforementioned policy probability distributions includes one or more parameters, The aforementioned learning policy determination means is: A second learning likelihood acquisition means that uses the quality information of the target image for training and each of the multiple policy probability distributions to determine the likelihood of each policy probability distribution, The system includes a third parameter updating means for updating the parameters of each policy probability distribution based on the first loss and a modified loss for modifying the parameters so that the difference between the parameters of the policy probability distributions corresponding to different network policies becomes larger, The third parameter update means is: A third loss acquisition means that obtains a third loss by adding the first loss and the corrected loss, Includes a third parameter determination means for updating the parameters of each policy probability distribution based on the third loss. The information processing system described in 11. 17. The system further comprises a learning policy determination means for training a third policy determination model for determining the network policy corresponding to the target image from among a predetermined number of network policies, The aforementioned learning policy determination means is: A means for obtaining a correct likelihood by inputting the quality information of the training target image into a trained first policy decision model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, and obtaining the likelihood of each network policy as the correct likelihood, A third learning likelihood acquisition means inputs the aforementioned training target images into a third policy decision model and obtains the likelihood of each network policy as the likelihood during training, The system includes a fourth parameter update means for updating the parameters of the third policy decision model based on the fourth loss such that the likelihood during learning approaches the true likelihood, The fourth loss is a loss based on the error between the likelihood during learning and the true likelihood. The information processing system described in 11. 18. Quality estimation means for estimating quality information of the target image, The system comprises an image processing means that processes the target image using a neural network that corresponds to a network policy determined based on the quality information of the target image. Information processing device. 19. One or more computers, Estimate the quality information of the target image, This includes processing the target image using a neural network that corresponds to a network policy determined based on the quality information of the target image. Information processing methods. 20. Processing the target image involves extracting features from the target image using the neural network according to the network policy. The information processing method described in 19. 21. The aforementioned target image is an eye image that includes the iris region in which the iris is reflected. The aforementioned quality information includes the iris diameter. The information processing method described in 19. or 20. 22. The quality information includes the resolution of the target image. An information processing method described in any one of items 19 to 21. 23. Further includes determining the network policy based on the quality information. An information processing method described in any one of items 19 to 22. 24. In determining the network policy, the network policy is determined such that the processing time using the neural network is within a predetermined range, regardless of the quality information of the target image. The information processing method described in 23. 25. Processing the aforementioned target image, The iris region included in the aforementioned target image is converted into a standard image of a predetermined shape. Using the neural network according to the aforementioned network policy, the standard image is processed, By converting to the aforementioned standard image, the iris region is converted to the aforementioned standard image of a size corresponding to the quality information. An information processing method described in any one of items 19 to 24. 26. The neural network is composed of a pre-stage group including an input layer and a post-stage group including a final layer. The aforementioned network policy is applied to image processing using the aforementioned group. An information processing method described in any one of items 19 to 25. 27. The neural network includes multiple convolutional layers, In determining the network policy, the network policy is determined such that the lower the resolution of the target image, the larger the number of convolutional layers included in the preceding group. The information processing method described in 26. 28. The neural network includes multiple convolutional layers, In determining the network policy, the lower the resolution of the target image, the larger the number of channels in the calculations performed in the convolutional layer included in the preceding group. The information processing method described in 26. 29. Further includes training the neural network by inputting training target images into the neural network. An information processing method described in any one of items 19 to 28. 30. Further comprising determining a network policy for the training target image based on the quality information and training probability distribution of the training target image, In training the neural network, the training of the neural network is performed by inputting the target images for training into the neural network according to the determined network policy. The aforementioned training probability distribution is one in which the highest probability is given for the quality indicated by the quality information of the training target image, and the probability gradually decreases as the degree of difference from that quality increases. The information processing method described in 29. 31. In determining the network policy for the aforementioned training images, By using the aforementioned random numbers, the quality information of the target image for training is modified according to the training probability distribution, and the training quality information is determined. Based on the determined training quality information, a network policy is determined for the training target images. The information processing method described in 30. 32. Further including selecting training images in such a way that lower-quality training images are selected more frequently, Training the neural network involves inputting the selected training target images into the neural network to perform the training. An information processing method described in any one of items 29 to 31. 33. Train a first policy decision model to determine the network policy corresponding to the quality information of the target image from among a predetermined number of network policies. The method further includes calculating the first loss when processing the training target image using a neural network corresponding to each of the aforementioned multiple network policies, The aforementioned first policy decision model consists of a neural network, In training the aforementioned first policy decision model, The quality information of the training target images is input into the first policy decision model to obtain the likelihood of each network policy. Based on the first loss and likelihood of each of the aforementioned network policies, the parameters of the first policy decision model are updated. Updating the parameters of the aforementioned first policy decision model means, For each of the aforementioned network policies, calculate the second loss, which is the sum of the first losses weighted by likelihood. Based on the second loss, update the parameters of the first policy decision model. The information processing method described in 29. 34. Train a second policy decision model to determine the network policy corresponding to the quality information of the target image from among a predetermined number of network policies. This further includes calculating the first loss when processing the training target image using a neural network corresponding to the network policy determined using the second policy determination model described above, The second policy decision model consists of a plurality of policy probability distributions associated with each of the plurality of network policies, Each of the aforementioned policy probability distributions includes one or more parameters, In training the aforementioned second policy decision model, Using the quality information of the training target images and each of the multiple policy probability distributions, the likelihood of each policy probability distribution is calculated. Based on the first loss and a modified loss for correcting the difference between the parameters of the policy probability distributions corresponding to different network policies, the parameters of each policy probability distribution are updated. By updating the parameters of each of the aforementioned policy probability distributions, The third loss is obtained by adding the first loss and the adjusted loss. Based on the third loss, update the parameters of each policy probability distribution. The information processing method described in 29. 35. Further includes training a third policy decision model for determining the network policy corresponding to the target image from among a predetermined number of network policies, In training the aforementioned third policy decision model, The quality information of the target image for training is input into a trained first policy decision model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, and the likelihood of each network policy is obtained as the correct likelihood. The aforementioned training target images are input into the third policy decision model, and the likelihood of each network policy is obtained as the likelihood during training. Based on the fourth loss, the parameters of the third policy decision model are updated so that the likelihood during learning approaches the true likelihood. The fourth loss is a loss based on the error between the likelihood during learning and the true likelihood. The information processing method described in 29. 36. On one or more computers, Estimate the quality information of the target image, A recording medium on which a program is stored that causes a neural network to process the target image using a network policy determined based on the quality information of the target image. 37. In processing the target image, a program for extracting features from the target image was recorded using the neural network according to the network policy. The recording medium described in 36. 38. The aforementioned target image is an eye image that includes the iris region in which the iris is reflected. The aforementioned quality information is recorded in a program that includes the iris diameter. The recording medium described in 36. or 37. 39. The quality information is recorded in a program that includes the resolution of the target image. A recording medium described in any one of items 36 to 38. 40. A program is recorded to further perform the task of determining the network policy based on the quality information. A recording medium described in any one of items 36 to 39. 41. In determining the network policy, a program for determining the network policy was recorded such that the processing time using the neural network is within a predetermined range, regardless of the quality information of the target image. The recording medium described in 40. 42. Processing the aforementioned target image, The iris region included in the aforementioned target image is converted into a standard image of a predetermined shape. Using the neural network according to the aforementioned network policy, the standard image is processed, In the process of converting to the aforementioned standard image, a program is recorded for converting the iris region into the aforementioned standard image of a size corresponding to the quality information. A recording medium described in any one of items 36 to 41. 43. The neural network is composed of a pre-stage group including an input layer and a post-stage group including a final layer. The aforementioned network policy contains a program for application to image processing using the aforementioned group. A recording medium described in any one of items 36 to 42. 44. The neural network includes multiple convolutional layers, In determining the aforementioned network policy, a program for determining the network policy was recorded such that the lower the resolution of the target image, the larger the number of convolutional layers included in the preceding group. The recording medium described in 43. 45. The neural network includes multiple convolutional layers, In determining the aforementioned network policy, a program for determining the network policy was recorded such that the lower the resolution of the target image, the larger the number of channels in the calculations performed in the convolutional layer included in the preceding group. The recording medium described in 43. 46. ​​A program is recorded that further includes training the neural network by inputting target images for training into the neural network. A recording medium described in any one of items 36 to 45. 47. Further comprising determining a network policy for the training target image based on the quality information and training probability distribution of the training target image, In training the neural network, the training of the neural network is performed by inputting the target images for training into the neural network according to the determined network policy. The program recorded is such that the training probability distribution has the highest probability at the quality indicated by the quality information of the training target image, and the probability gradually decreases as the degree of difference from that quality increases. The recording medium described in 46. 48. In determining the network policy for the aforementioned training images, By using the aforementioned random numbers, the quality information of the target image for training is modified according to the training probability distribution, and the training quality information is determined. A program for determining a network policy for the target images to be trained, based on the determined training quality information, was recorded. The recording medium described in 47. 49. Further including selecting training images in such a way that lower-quality training images are selected more frequently, In training the neural network, the selected training target images are input to the neural network, and a program for training the neural network is recorded. A recording medium described in any one of items 46 to 48. 50. A first policy decision model is trained to determine the network policy corresponding to the quality information of the target image from among a predetermined number of network policies. The method further includes calculating the first loss when processing the training target image using a neural network corresponding to each of the aforementioned multiple network policies, The aforementioned first policy decision model consists of a neural network, In training the aforementioned first policy decision model, The quality information of the training target images is input into the first policy decision model to obtain the likelihood of each network policy. Based on the first loss and likelihood of each of the aforementioned network policies, the parameters of the first policy decision model are updated. Updating the parameters of the aforementioned first policy decision model means, For each of the aforementioned network policies, calculate the second loss, which is the sum of the first losses weighted by likelihood. Based on the second loss, a program was recorded to update the parameters of the first policy decision model. The recording medium described in 46. 51. Train a second policy decision model to determine the network policy corresponding to the quality information of the target image from among a predetermined number of network policies. This further includes calculating the first loss when processing the training target image using a neural network corresponding to the network policy determined using the second policy determination model described above, The second policy decision model consists of a plurality of policy probability distributions associated with each of the plurality of network policies, Each of the aforementioned policy probability distributions includes one or more parameters, In training the aforementioned second policy decision model, Using the quality information of the training target images and each of the multiple policy probability distributions, the likelihood of each policy probability distribution is calculated. Based on the first loss and a modified loss for correcting the difference between the parameters of the policy probability distributions corresponding to different network policies, the parameters of each policy probability distribution are updated. By updating the parameters of each of the aforementioned policy probability distributions, The third loss is obtained by adding the first loss and the adjusted loss. Based on the third loss, a program was recorded for updating the parameters of each policy probability distribution. The recording medium described in 46. 52. Further includes training a third policy decision model for determining the network policy corresponding to the target image from among a predetermined number of network policies, In training the aforementioned third policy decision model, The quality information of the target image for training is input into a trained first policy decision model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, and the likelihood of each network policy is obtained as the correct likelihood. The aforementioned training target images are input into the third policy decision model, and the likelihood of each network policy is obtained as the likelihood during training. Based on the fourth loss, the parameters of the third policy decision model are updated so that the likelihood during learning approaches the true likelihood. A program was recorded in which the fourth loss is a loss based on the error between the likelihood during learning and the true likelihood. The recording medium described in 46.

[0436] This application claims priority based on PCT / JP2022 / 041157, filed on 4 November 2022, and incorporates all of its disclosures herein. [Explanation of Symbols]

[0437] 100, 400, 500, 600 Information Processing Systems 101 Imaging device 102,403,503,603 Information Processing Equipment 111 Quality estimation section 112 Policy Decision-Making Department 113 Image Processing Unit 113a First Processing Unit 113b Second Processing Unit 114 Authentication Department 421,521 Learning memory section 422 Learning Department 423 Learning Policy Determination Unit 423a 1st decision section 423b Second decision section 524 Image Selection Section 525 Learning Policy Decision Unit NN1~NN3 Neural Networks

Claims

1. A quality estimation means for estimating the quality information of the target image, Image processing means that processes the target image using a neural network according to a network policy determined based on the quality information of the target image, A learning means that trains the neural network by inputting training target images into the neural network, A learning policy determination means for training a first policy determination model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, The system includes a first loss acquisition means for determining the first loss when processing the training target image using a neural network corresponding to each of the aforementioned multiple network policies, The aforementioned first policy decision model consists of a neural network, The aforementioned learning policy determination means is: A first learning likelihood acquisition means inputs the quality information of the target images for training into the first policy decision model and calculates the likelihood of each network policy, The system includes a second parameter updating means for updating the parameters of the first policy decision model based on the first loss and likelihood of each of the network policies, The second parameter updating means is A second loss acquisition means for obtaining a second loss which is the sum of the first losses weighted by likelihood for each of the aforementioned network policies, Includes a second parameter update means for updating the parameters of the first policy decision model based on the second loss. Information processing system.

2. A quality estimation means for estimating the quality information of the target image, Image processing means that processes the target image using a neural network according to a network policy determined based on the quality information of the target image, A learning means that trains the neural network by inputting training target images into the neural network, The system includes a learning policy determination means for training a third policy determination model for determining the network policy corresponding to the target image from among a predetermined number of network policies, The aforementioned learning policy determination means is: A means for obtaining a true likelihood by inputting the quality information of the training target image into a trained first policy decision model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, and obtaining the likelihood of each network policy as the true likelihood, A third learning likelihood acquisition means inputs the aforementioned training target images into a third policy decision model and obtains the likelihood of each network policy as the likelihood during training, The system includes a fourth parameter update means for updating the parameters of the third policy decision model based on the fourth loss, such that the likelihood during learning approaches the true likelihood, The fourth loss is a loss based on the error between the likelihood during learning and the true likelihood. Information processing system.

3. The image processing means extracts features from the target image using the neural network according to the network policy. The information processing system according to claim 1 or 2.

4. The aforementioned target image is an eye image that includes the iris region in which the iris is reflected. The aforementioned quality information includes the iris diameter. The information processing system according to claim 1 or 2.

5. One or more computers, Estimate the quality information of the target image, The target image is processed using a neural network that corresponds to a network policy determined based on the quality information of the target image. By inputting the training target images into the neural network, the neural network is trained. A first policy decision model is trained to determine the network policy corresponding to the quality information of the target image from among a predetermined number of network policies. This includes calculating the first loss when processing the training target image using a neural network corresponding to each of the aforementioned multiple network policies, The aforementioned first policy decision model consists of a neural network, Training the aforementioned first policy decision model is The quality information of the training target images is input into the first policy decision model to obtain the likelihood of each network policy. This includes updating the parameters of the first policy decision model based on the first loss and likelihood of each of the aforementioned network policies, Updating the parameters of the first policy decision model mentioned above means For each of the aforementioned network policies, calculate the second loss, which is the sum of the first losses weighted by likelihood. This includes updating the parameters of the first policy decision model based on the second loss. Information processing methods.

6. One or more computers, Estimate the quality information of the target image, The target image is processed using a neural network that corresponds to a network policy determined based on the quality information of the target image. By inputting the training target images into the neural network, the neural network is trained. This includes training a third policy decision model for determining the network policy corresponding to the target image from among a predetermined number of network policies, Training the aforementioned third policy decision model is The quality information of the target image for training is input into a trained first policy decision model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, and the likelihood of each network policy is obtained as the correct likelihood. The aforementioned training target images are input into the third policy decision model, and the likelihood of each network policy is obtained as the likelihood during training. This includes updating the parameters of the third policy decision model based on the fourth loss so that the likelihood during learning approaches the true likelihood, The fourth loss is a loss based on the error between the likelihood during learning and the true likelihood. Information processing methods.

7. On one or more computers, Estimate the quality information of the target image, The target image is processed using a neural network that corresponds to a network policy determined based on the quality information of the target image. By inputting the training target images into the neural network, the neural network is trained. A first policy decision model is trained to determine the network policy corresponding to the quality information of the target image from among a predetermined number of network policies. The process involves calculating the first loss when processing the target image for training using a neural network corresponding to each of the aforementioned multiple network policies. The aforementioned first policy decision model consists of a neural network, Training the aforementioned first policy decision model is The quality information of the training target images is input into the first policy decision model to obtain the likelihood of each network policy. This includes updating the parameters of the first policy decision model based on the first loss and likelihood of each of the aforementioned network policies, Updating the parameters of the first policy decision model mentioned above means For each of the aforementioned network policies, calculate the second loss, which is the sum of the first losses weighted by likelihood. A program that includes updating the parameters of the first policy decision model based on the second loss.

8. On one or more computers, Estimate the quality information of the target image, The target image is processed using a neural network that corresponds to a network policy determined based on the quality information of the target image. By inputting the training target images into the neural network, the neural network is trained. The system is made to perform training on a third policy decision model for determining the network policy corresponding to the target image from among a predetermined number of network policies. Training the aforementioned third policy decision model is The quality information of the target image for training is input into a trained first policy decision model for determining the network policy corresponding to the quality information of the target image from among a predetermined number of network policies, and the likelihood of each network policy is obtained as the correct likelihood. The aforementioned training target images are input into the third policy decision model, and the likelihood of each network policy is obtained as the likelihood during training. This includes updating the parameters of the third policy decision model based on the fourth loss so that the likelihood during learning approaches the true likelihood, The fourth loss is a loss based on the error between the likelihood during learning and the correct likelihood in the program.

Images