Learning device, learning method, determination device, determination method and computer program

Abstract

To provide a learning device, a learning method, a determination device, a determination method, and a computer program that enable more accurately determining subjects of an image using spectral images.SOLUTION: In a determination system including a learning device and a determination device connected to an imaging device for capturing multispectral images, the learning device 20 comprises a control unit 23 which uses multispectral images obtained by capturing each of a plurality of subjects to generate a feature extraction model for obtaining feature amounts from the multispectral images using a first loss function that results in a smaller value of a feature amount obtained from a multispectral image of an identical subject, and a second loss function that results in a value that is small enough that a prescribed disturbance can be more accurately identified, so that the first loss function has a smaller value and the second loss function has a larger value.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a learning device, a learning method, a determination device, a determination method, and a computer program. [Background technology]

[0002] As a technology for non-contact personal authentication from spectral images and image sequences including RGB images, technologies that use characteristic facial patterns, such as face recognition and iris recognition, are common. However, this technology has a problem of high psychological resistance to photography. Furthermore, it has significant restrictions, such as being limited to an angle of view that captures the face. Therefore, there is a technology that uses an image of a hand as a non-contact authentication method that is less likely to interfere with the freedom of the person to be authenticated. For example, there is a technology that uses the width and height of the palm and fingers, and a technology that uses the outline shape of the hand. However, these technologies have restrictions, such as the palm being fixed to a stand and photographed in order to extract the shape accurately. There is also a technology that authenticates individuals based on hand gestures. However, it is necessary to request a specific action from the person to be authenticated. Technologies that use multispectral images for personal authentication have also been proposed. For example, in Non-Patent Document 1, a multispectral image of a hand is used to perform personal authentication without using a specific movement or restraining the subject. The technology in Non-Patent Document 1 uses a multispectral image to improve accuracy to a certain extent compared to RGB images. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Takeshi Uemori, Atsushi Ito, Yusuke Moriuchi, Alexander Gatto, and Jun Murayama.”Skin-based identification from multispectral image data using CNNs.”, In Proc. CVPR, 2019. Summary of the Invention [Problem to be solved by the invention]

[0004] As mentioned above, the method using spectral images can realize personal authentication from images of the back of a hand or other images with few features without restricting the freedom of the subject. On the other hand, because the spectral resolution is high, it is vulnerable to disturbances such as hand posture (whether the hand is clenched or closed) and temperature changes. Therefore, the same person may be recognized as a different person if the hand posture or temperature is different. Conversely, different people may be recognized as the same person. The present invention has been made in consideration of the above-mentioned circumstances, and provides a technique that enables a subject of an image to be determined with higher accuracy using a spectral image. [Means for solving the problem]

[0005] One aspect of the present invention is a learning device that includes a control unit that uses multispectral images obtained by photographing each of a plurality of subjects, and generates a feature extraction model for acquiring feature amounts from the multispectral images using a first loss function such that a feature amount obtained from a multispectral image of the same subject has a smaller value, and a second loss function such that a value is small enough to more accurately identify a predetermined disturbance, such that the first loss function has a smaller value and the second loss function has a larger value.

[0006] One aspect of the present invention is a learning method that uses multispectral images obtained by photographing each of a plurality of subjects, and generates a feature extraction model for acquiring feature quantities from the multispectral images using a first loss function that results in a smaller value of a feature quantity obtained from a multispectral image of the same subject, and a second loss function that results in a value that is small enough that a predetermined disturbance can be more accurately identified, so that the first loss function has a smaller value and the second loss function has a larger value.

[0007] One aspect of the present invention is a determination device comprising: an image input unit that acquires a multispectral image of a target object, and a control unit that acquires feature amounts from the multispectral image of the target object using multispectral images obtained by photographing each of a plurality of targets, the feature amounts being smaller than a value of a feature amount obtained from a multispectral image of the same target object and a second loss function being smaller in value so that a predetermined disturbance can be more accurately identified, the control unit acquiring feature amounts from the multispectral image of the target object using a feature extraction model that acquires feature amounts from the multispectral image so that the first loss function has a smaller value and the second loss function has a larger value, and that determines the target object of the multispectral image of the target object based on the acquired feature amounts.

[0008] One aspect of the present invention is a determination method comprising: acquiring a multispectral image of a target object; and using multispectral images obtained by photographing each of a plurality of subjects, acquiring feature amounts from the multispectral image of the target object using a feature extraction model that acquires feature amounts from the multispectral image using a first loss function such that a feature amount obtained from a multispectral image of the same subject has a smaller value, and a second loss function such that a value is small enough that a predetermined disturbance can be more accurately identified, so that the first loss function has a smaller value and the second loss function has a larger value; and determining the target object of the multispectral image of the target object based on the acquired feature amounts.

[0009] One aspect of the present invention is a computer program for causing a computer to function as the learning device or the determination device described above. Effect of the Invention

[0010] According to the present invention, it is possible to more accurately determine the subject of an image using a spectral image. [Brief description of the drawings]

[0011] [Figure 1] 1 is a schematic block diagram showing a system configuration of a determination system 10 according to the present invention. [Diagram 2] 2 is a schematic block diagram showing a specific example of the functional configuration of a learning device 20. FIG. [Diagram 3] 4 is a flowchart showing a specific example of processing by the learning device 20. [Figure 4] 2 is a schematic block diagram showing a specific example of the functional configuration of a determination device 30. FIG. [Diagram 5] 4 is a flowchart showing a specific example of the process of the determination device 30. [Figure 6] FIG. 2 is a diagram illustrating an outline of a hardware configuration example of an information processing device 90 applied to the present embodiment. [Figure 7] FIG. 13 is a diagram showing a modified example of the determination device 30. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] FIG. 1 is a schematic block diagram showing a system configuration of a determination system 10 of the present invention. The determination system 10 is used when determining a determination object based on a multispectral image (hereinafter referred to as a "target multispectral image") obtained by capturing an image of the determination object. The determination object may be, for example, a living object or a non-living object. When the determination object is a non-living object, the determination object may be, for example, a material such as a metal or a polymer compound, or may be a tangible object such as a desk or a car. When the determination object is a living object, the determination object may be, for example, an animal, a plant, a fungus, or a virus. When the determination object is an animal, the determination object may be a human, a mammal, or another animal. When the determination object is a human, it is desirable that the target multispectral image is an image of a predetermined specific part of the human. For example, the target multispectral image may be an image of a part or the whole of a human body such as a human palm or a human face.

[0013] The determination using the target multispectral image in the determination system 10 may be, for example, a determination of the type of the determination target captured in the target multispectral image (e.g., a type of animal, a type of plant, a type of fungus, etc.). The determination using the target multispectral image may be a determination of the attributes (e.g., age, sex, etc.) of a specific organism (e.g., a person), or a determination of an individual organism (e.g., a person) (e.g., personal authentication, etc.).

[0014] The determination system 10 includes a learning device 20 and a determination device 30. An imaging device 40 that captures a multispectral image is connected to the determination device 30. The "multispectral image" in this embodiment is an image having a plurality of wavelength information of at least four bands or more, and does not need to be limited to a specific numerical value of four or more. The imaging device 40 captures an image with a higher spectral resolution than a camera that captures at least RGB. For example, a so-called multispectral camera (having about 10 wavelengths) or a hyperspectral camera (having 10 or more to several hundred wavelengths) may be used as the imaging device 40. The multispectral image used in the determination system 10 is an image captured by such a multispectral camera or hyperspectral camera.

[0015] The learning device 20 and the determination device 30 may be communicably connected via a network. The network may be a network using wireless communication or a network using wired communication. The network may be configured using, for example, the Internet or a local area network (LAN). The network may be configured by combining a plurality of networks.

[0016] 2 is a schematic block diagram showing a specific example of the functional configuration of learning device 20. Learning device 20 is configured using an information processing device such as a personal computer or a server device. Learning device 20 includes a communication unit 21, a storage unit 22, and a control unit 23.

[0017] The communication unit 21 is a communication device. The communication unit 21 may be configured as, for example, a network interface. The communication unit 21 performs data communication with other devices via a network in response to control by the control unit 23. The communication unit 21 may be a device that performs wireless communication or a device that performs wired communication.

[0018] The storage unit 22 is configured using a storage device such as a magnetic hard disk device or a semiconductor storage device. The storage unit 22 stores data used by the control unit 23. The storage unit 22 may function as, for example, a basic data storage unit 221 and a feature extraction model storage unit 222.

[0019] The basic data storage unit 221 stores basic data used in the learning process executed in the learning device 20. The basic data includes data on a plurality of multispectral images obtained by capturing images of a plurality of objects to be determined. The basic data further includes label information indicating the correct answer of information to be obtained as a determination result of the object to be determined captured in each multispectral image. For example, in the case of a determination system 10 in which the type of the object to be determined is determined as the determination result, the label information indicates the correct answer of the type of the object to be determined. For example, in the case of a determination system 10 in which the individual of the object to be determined is determined as the determination result and an authentication result is obtained, the label information indicates the correct answer of the individual of the object to be determined (person).

[0020] The feature extraction model storage unit 222 stores a feature extraction model obtained by executing a learning process by the control unit 23. The feature extraction model indicates a model for acquiring a feature amount from a multispectral image. The feature extraction model will be described in detail later.

[0021] The control unit 23 is configured using a processor such as a CPU (Central Processing Unit) and a memory (main storage device). The control unit 23 functions by the processor executing a program. All or part of the functions of the control unit 23 may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). The above program may be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or a semiconductor storage device (e.g., an SSD: Solid State Drive), or a storage device such as a hard disk or a semiconductor storage device built into a computer system. The above program may be transmitted via an electric communication line.

[0022] The information control unit 231 controls the input and output of information. For example, the information control unit 231 acquires basic data from other devices (information processing devices and storage media) and records the basic data in the basic data storage unit 221. For example, the information control unit 231 transmits the feature extraction model stored in the feature extraction model storage unit 222 to other devices (for example, the determination device 30). For example, the information control unit 231 outputs the feature extraction model stored in the feature extraction model storage unit 222 to other devices (for example, a recording medium).

[0023] The learning control unit 232 executes a learning process using the basic data stored in the basic data storage unit 221. The learning control unit 232 acquires a feature extraction model as a result of the learning process.

[0024] The details of the processing of the learning control unit 232 will be described. The learning control unit 232 learns a feature extraction model (feature extractor) using a group of multispectral images and a group of label information included in the basic data as input. In the learning processing of the learning control unit 232, a disturbance category is explicitly introduced into the framework of supervised control learning in order to learn features that have high individual authentication accuracy and are invariant to specified disturbances. Specifically, the learning control unit 232 uses a first loss function and a second loss function.

[0025] The first loss function has a smaller value as the features are closer to each other, such as between multispectral images with the same ground truth label, and a larger value as the features are farther apart, such as between multispectral images with different ground truth labels. A specific example of such a first loss function is contrastive loss. For example, the loss of Supervised Contrastive Learning may be used as the first loss function.

[0026] The second loss function has a smaller value as the specified disturbance can be more accurately identified. A specific example of such a second loss function is identification loss. For example, cross entropy may be used as the second loss function. The learning control unit 232 learns the feature extraction model so that the value of the first loss function becomes smaller and the value of the second loss function becomes larger.

[0027] A more specific example of the processing of the learning control unit 232 will be described. In the following description, I indicates a set of indexes of multispectral images. i is an arbitrary value included in I. x_i indicates a multispectral image of index i. Note that the description "A_B" indicates that a subscript "B" is attached to a certain letter "A". P(i) indicates a set of indexes of multispectral images of the same person as the multispectral image x_i. y_i is defined as follows and is a one-hot vector indicating the disturbance category of the multispectral image x_i.

[0028]

number

[0029] c indicates the number of categories. The categories may be defined in any way as long as they are given and categorical. For example, a category may be defined that indicates the time of day (morning, daytime, night, etc.) when the multispectral image was taken, a category may be defined that indicates the state of the object to be determined (for example, rock, scissors, paper, etc. for a hand), or a type of camera (for example, camera model number). The category (disturbance category) is used as a correct answer value when performing discrimination learning. By using the disturbance category in this way, it is possible to learn a feature extractor in which data of the same object to be determined is closely mapped and the disturbance category cannot be classified.

[0030] G indicates a feature extraction model (feature extractor). For example, a feature extraction backbone of ResNet may be applied to the feature extraction model. F indicates a disturbance discrimination head. For example, a discrimination head of ResNet may be applied. In addition, softmax may be applied as an activation function, or softmax with temperature may be used. L_con indicates a first loss function (for example, contrastive loss). For example, a Supervised Contrastive Learning loss may be applied. L_cls indicates a second loss function (for example, discrimination loss). For example, Cross Entropy may be applied.

[0031] The learning control unit 232 may perform the learning process by repeatedly executing contrast learning and discriminatory learning as shown below. Note that the repetition may involve, for example, performing contrast learning once, then discriminatory learning once, returning to contrast learning once, performing discriminatory learning once, and then returning to contrast learning. As another embodiment, contrast learning may be performed multiple times, and then discriminatory learning may be repeated multiple times to complete the process. First, we will explain the contrastive learning. In contrastive learning, we first calculate z_i for any i using the following formula. This calculation extracts the features of the multispectral image x_i.

[0032]

number

[0033] Next, the learning control unit 232 calculates the contrast loss using the following formula. In the following formula, τ is a hyperparameter that is adjusted during learning using temperature softmax. Also, the following formula (3) is merely a specific example, and any loss function used in so-called contrast learning may be used.

[0034]

number

[0035] Then, the learning control unit 232 updates the parameters of the feature extraction model G so as to minimize the calculated contrast loss L_con. Specifically, the parameters are updated in the negative gradient direction of L_con. This completes the contrast learning process.

[0036] Next, discriminative learning (particularly, adversarial discriminative learning) will be described. In discriminative learning, first, z_i is calculated for any i using the above formula. This calculation extracts the features of the multispectral image x_i. Next, the learning control unit 232 executes disturbance discrimination processing for z_i. Specifically, ^y_i is calculated using the following formula. Note that the description "^A" indicates that a "^" is placed above a certain letter "A". The function F in the following formula (4) is a disturbance discrimination function, and for example, the output layer of ResNet can be used. In addition, any function used in a general discrimination model may be used for the function F.

[0037]

number

[0038] Next, the learning control unit 232 calculates the classification loss using the following formula: In the following formula, y corresponds to the above-mentioned disturbance category.

number

[0039] Then, the learning control unit 232 updates the parameters of the disturbance identification head F so as to minimize the calculated identification loss L_cls. Specifically, the parameters are updated in the negative gradient direction of L_cls. Furthermore, the learning control unit 232 updates the parameters of the feature extraction model G so as to maximize the identification loss L_cls. Specifically, the parameters are updated in the positive gradient direction of L_cls. In other words, first, the disturbance identification head F is trained so that the disturbance category can be classified with the feature extraction model G fixed. Then, the disturbance identification head F is trained so that the disturbance category cannot be classified with the disturbance identification head F fixed. The above is the process of adversarial identification learning. The learning control unit 232 generates a feature extraction model by executing the above process. The learning control unit 232 records the generated feature extraction model in the feature extraction model storage unit 222.

[0040] 3 is a flowchart showing a specific example of processing of the learning device 20. First, the information control unit 231 acquires basic data (step S101). The basic data may be input by an administrator of the learning device 20, may be acquired from another information device through communication, or may be acquired from a recording medium connected to the learning device 20. The learning control unit 232 performs a learning process by executing a comparison process and adversarial discrimination learning using the basic data. The learning control unit 232 records a feature extraction model generated by the learning process in the feature extraction model storage unit 222 (step S102).

[0041] 4 is a schematic block diagram showing a specific example of a functional configuration of the determination device 30. The determination device 30 is configured using an information processing device such as a personal computer or a server device. The determination device 30 includes an image input unit 31, a storage unit 32, a control unit 33, and an output unit .

[0042] The image input unit 31 accepts multispectral image data input to the determination device 30. The image input unit 31 accepts input of multispectral image data captured by the imaging device 40. There is no limitation on the manner in which the image input unit 31 accepts input of multispectral image data from the imaging device 40. The image input unit 31 may be connected to the imaging device 40 via, for example, a USB (Universal Serial Bus) or other communication cable, and may accept input of image data via the communication cable. The image input unit 31 may be communicably connected to the imaging device 40 via wireless communication such as Bluetooth (registered trademark), and may accept input of image data.

[0043] The storage unit 32 is configured using a storage device such as a magnetic hard disk device or a semiconductor storage device. The storage unit 32 stores data used by the control unit 33. The storage unit 32 may function as, for example, a feature extraction model storage unit 321 and a determination information storage unit 322.

[0044] The feature extraction model storage unit 321 stores a feature extraction model used when the feature extraction unit 332 performs feature extraction processing. The feature extraction model is generated by the learning device 20 in advance.

[0045] The determination information storage unit 322 stores the feature amounts extracted from the multispectral image by the feature extraction unit 332 in association with information indicating the subject having the feature amount (the subject of the multispectral image). The information indicating the subject is, for example, information equivalent to a correct answer label.

[0046] The control unit 33 is configured using a processor such as a CPU and a memory. The control unit 33 functions as an information control unit 331, a feature extraction unit 332, and a determination unit 333 by the processor executing a program. All or part of the functions of the control unit 33 may be realized using hardware such as an ASIC, a PLD, or an FPGA. The above program may be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or a semiconductor storage device (e.g., SSD), or a storage device such as a hard disk or a semiconductor storage device built into a computer system. The above program may be transmitted via an electric communication line.

[0047] The information control unit 331 acquires a feature extraction model from another device such as the learning device 20. The information control unit 331 acquires multispectral image data input from the image input unit 31. The information control unit 331 outputs information indicating the determination result obtained by the determination unit 333 to the outside via the output unit 34. For example, the information control unit 331 may transmit the information indicating the determination result to another device, or may output the information indicating the determination result from the output unit 34 via a display or speaker.

[0048] The feature extraction unit 332 acquires a feature amount by using a feature extraction model for the multispectral image input from the image input unit 31. That is, the feature extraction unit 332 acquires a feature amount z=G(x) for the multispectral image x to be determined that is input from the image input unit 31 by using the feature extraction model G.

[0049] The determination unit 333 determines the subject of the multispectral image x to be determined based on the feature amount acquired by the feature extraction unit 332 and the information stored in the determination information storage unit 322. For example, when the determination device 30 is a device for authenticating an individual, information indicating an individual (personal identification information) and information on the feature amount are stored in the determination information storage unit 322. The determination unit 333 determines which information on the feature amount stored in the determination information storage unit 322 matches the feature amount acquired from the multispectral image x to be determined, and acquires the personal identification information on the subject. For example, the determination unit 333 may select the feature amount closest to the feature amount acquired by the feature extraction unit 332 from among the feature amounts recorded in the determination information storage unit 322, and determine that the determination subject is associated with the feature amount if the difference (distance) is within a predetermined threshold value.

[0050] 5 is a flowchart showing a specific example of the process of the determination device 30. First, the information control unit 331 acquires a multispectral image of the determination target captured by the imaging device 40 (step S201). The feature extraction unit 332 acquires a feature amount of the multispectral image of the determination target using a feature extraction model stored in the feature extraction model storage unit 321 (step S202). The determination unit 333 performs a determination process on the subject of the multispectral image of the determination target using the information stored in the determination information storage unit 322 and the feature amount acquired by the feature extraction unit 332 (step S203). The determination unit 333 outputs information indicating the determination result via the output unit 34 (step S204).

[0051] The determination system 10 configured in this way can realize a determination (e.g., personal authentication) that is robust against the influence of disturbances from a spectral image of a determination target object with few features (e.g., the back of a person's hand) without contacting the subject and without placing a load on the subject. This allows for flexible determination (e.g., personal authentication) to be realized depending on the installation position of the camera and the movement of the subject.

[0052] Specifically, it is as follows. The judgment system 10 learns features that are invariant (with little change) to a specified disturbance. In this way, a disturbance category is explicitly introduced into the framework of supervised control learning. By specifying a disturbance, the effect of the disturbance can be suppressed, enabling highly accurate judgment (e.g., personal authentication). A variable indicating the disturbance category mixed into each spectral image is input, and features that are invariant to the disturbance are learned.

[0053] In addition, the contrastive learning and the adversarial low discriminative learning are independent. Therefore, it is possible to generate a feature extraction model even if y_i is not given for all x_i.

[0054] FIG. 6 is a diagram showing an outline of a hardware configuration example of an information processing device 90 applied to this embodiment. The information processing device 90 includes a processor 91, a main storage device 92, a communication interface 93, an auxiliary storage device 94, an input / output interface 95, and an internal bus 96. The processor 91, the main storage device 92, the communication interface 93, the auxiliary storage device 94, and the input / output interface 95 are communicably connected to each other via the internal bus 96. The information processing device 90 may be applied to, for example, the learning device 20 and the determination device 30. In this case, for example, the communication unit 21 and the output unit 34 may be configured using the communication interface 93. For example, the image input unit 31 and the output unit 34 may be configured using the input / output interface 95. For example, the memory unit 22 and the memory unit 32 may be configured using the auxiliary storage device 94. In addition, the control unit 23 and the control unit 33 may be configured using the processor 91 and the main storage device 92.

[0055] (Modification) In this embodiment, the learning device 20 and the determination device 30 are configured as different devices, but they may be configured as an integrated device. FIG. 7 is a diagram showing a modified example of the determination device 30 configured in this manner. The memory unit 32 of the determination device 30 shown in FIG. 7 also functions as a basic data memory unit 323. The control unit 33 of the determination device 30 shown in FIG. 7 also functions as a learning control unit 334. The basic data memory unit 323 and the learning control unit 334 function in the same way as the basic data memory unit 221 and the learning control unit 232 of the learning device 20, respectively.

[0056] The learning device 20 may be implemented using a plurality of information processing devices. For example, the learning device 20 may be implemented using a device such as a cloud. For example, in the learning device 20, the memory unit 22 and the control unit 23 may be implemented in different information processing devices. For example, the memory unit 22 of the learning device 20 may be distributed and implemented in a plurality of information processing devices. The determination device 30 may be implemented using a plurality of information processing devices. For example, the determination device 30 may be implemented using a device such as a cloud. For example, in the determination device 30, the memory unit 32 and the control unit 33 may be implemented in different information processing devices. For example, the memory unit 32 of the determination device 30 may be distributed and implemented in a plurality of information processing devices.

[0057] The learning control unit 232 may operate as follows. First, the learning control unit 232 extracts features from the input multispectral image using equation (2). Next, the learning control unit 232 calculates ^y_i, which is the disturbance identification result, using equation (4). Next, the learning control unit 232 calculates the contrast loss using equation (3). Next, the learning control unit 232 calculates the identification loss using equation (5). Next, the learning control unit 232 updates the parameters of the disturbance identification head F so as to minimize the calculated identification loss L_cls. Then, the learning control unit 232 performs feature learning. Specifically, the learning control unit 232 updates the parameters of the feature extraction model G so as to minimize L_con-λL_cls. Specifically, the parameters are updated in the negative gradient direction of L_con-λL_cls. The above is the process of contrast learning. By being configured in this way, the balance between contrast learning and adversarial low discrimination learning can be explicitly controlled by the hyperparameter λ. Therefore, an effect is obtained in which it is easy to obtain features with higher accuracy.

[0058] Although an embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and designs that do not deviate from the gist of the present invention are also included. [Explanation of symbols]

[0059] 10...judgment system, 20...learning device, 21...communication unit, 22...storage unit, 221...basic data storage unit, 222...feature extraction model storage unit, 23...control unit, 231...information control unit, 232...learning control unit, 30...judgment device, 31...image input unit, 32...storage unit, 321...feature extraction model storage unit, 322...judgment information storage unit, 33...control unit, 331...information control unit, 332...feature extraction unit, 333...judgment unit, 34...output unit, 40...imaging device

Claims

1. a control unit that uses multispectral images obtained by photographing each of a plurality of subjects, and generates a feature extraction model for acquiring feature amounts from the multispectral images using a first loss function such that a feature amount obtained from a multispectral image of a same subject has a smaller value, and a second loss function such that a value is small enough to more accurately identify a predetermined disturbance, so that the first loss function has a smaller value and the second loss function has a larger value.

2. A learning method using multispectral images obtained by photographing each of a plurality of subjects, the method using a first loss function such that a feature amount obtained from a multispectral image of the same subject has a smaller value, and a second loss function such that a value is small enough to more accurately identify a predetermined disturbance, to generate a feature extraction model for acquiring feature amounts from the multispectral images such that the first loss function has a smaller value and the second loss function has a larger value.

3. an image input unit for acquiring a multispectral image of an object to be determined; a control unit that uses multispectral images obtained by photographing each of a plurality of subjects, and acquires feature amounts from the multispectral images to be determined using a feature extraction model that acquires feature amounts from the multispectral images using a first loss function such that a feature amount obtained from a multispectral image of the same subject has a smaller value, and a second loss function such that a value is small enough to more accurately identify a predetermined disturbance, so that the first loss function has a smaller value and the second loss function has a larger value, and that determines the subject of the multispectral image to be determined based on the acquired feature amounts; A determination device comprising:

4. Acquire a multispectral image of the object to be judged; A determination method comprising: using multispectral images obtained by photographing each of a plurality of subjects, acquiring feature amounts from the multispectral images to be determined based on the acquired feature amounts using a feature extraction model that acquires feature amounts from the multispectral images using a first loss function such that a feature amount obtained from a multispectral image of the same subject has a smaller value, and a second loss function such that a value is small enough to more accurately identify a predetermined disturbance, so that the first loss function has a smaller value and the second loss function has a larger value; and determining the subject of the multispectral image to be determined based on the acquired feature amounts.

5. A computer program for causing a computer to function as the learning device according to claim 1 or the determination device according to claim 3.

Images