Learning system, estimation system, learning method, and computer program

Abstract

A learning system (10) according to the present invention comprises: a generation means (110) that combines an original image including a living body with information pertaining to a light reflection component, which is not derived from the original image, to generate a composite image; and a training means (130) that trains an inference model (121) on the basis of information pertaining to a particular site of the living body which has been inferred from the composite image by the inference model and correct answer information which indicates the correct answer of information pertaining to the particular site in the original image. The learning system makes it possible to appropriately train the inference model.

Description

Technical Field

[0001] This disclosure relates to the technical fields of a learning system, an estimation system, a learning method, and a computer program that execute learning of an estimation model.

Background Art

[0002] As this type of system, there is known one that executes processing related to a light reflection component in an image (i.e., a region where reflected light is reflected). For example, Patent Document 1 discloses a technique of executing a compositing process of specular and diffuse reflection components after executing a scaling process of a separated specular reflection component. Patent Document 2 discloses a technique of separating a specular reflection component, performing image conversion only on a diffuse reflection image, and recombining it with the specular reflection image.

[0003] As other related techniques, for example, Patent Document 3 discloses a technique of updating parameters of a machine learning model based on a gradient vector calculated from the difference between estimated data and correct data. Patent Document 4 discloses a technique of learning a component recognition classifier used for face recognition by machine learning.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Patent Document 3

Patent Document 4

Summary of the Invention

Problems to be Solved by the Invention

[0005] This disclosure has been made, for example, in view of the above-mentioned respective cited documents, and an object thereof is to provide a learning system, an estimation system, a learning method, and a computer program capable of appropriately learning an estimation model for a living body.

Means for Solving the Problem

[0006] One aspect of the learning system of this disclosure includes a generation means for generating a synthetic image by synthesizing information on a reflection component of light not derived from the original image with the original image including a living body, and a learning means for learning the estimation model based on information on a specific part of the living body estimated from the synthetic image by the estimation model and correct answer information indicating the correct answer of information on the specific part in the original image.

[0007] One aspect of the estimation system of this disclosure includes a generation means for generating a synthetic image by synthesizing information on a reflection component of light not derived from the original image with the original image including a living body, a learning means for learning the estimation model based on information on a specific part of the living body estimated from the synthetic image by the estimation model and correct answer information indicating the correct answer of information on the specific part in the original image, the estimation model learned using the above, and an output means for estimating and outputting information on the specific part in the input image using the estimation model.

[0008] One aspect of the learning method of this disclosure is to generate a synthetic image by synthesizing information on a reflection component of light not derived from the original image with the original image including a living body, and to perform learning of the estimation model based on information on a specific part of the living body estimated from the synthetic image by the estimation model and correct answer information indicating the correct answer of information on the specific part in the original image.

[0009] One aspect of the computer program of this disclosure is to generate a synthetic image by synthesizing information regarding the reflection component of light not derived from the original image into the original image including a living body, and to cause a computer to operate so as to perform learning of the estimation model based on information regarding a specific part of the living body estimated from the synthetic image by the estimation model and correct answer information indicating the correct answer of the information regarding the specific part in the original image.

Brief Description of the Drawings

[0010]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Figure 16

Figure 17

Figure 18

Figure 19

Figure 20

Figure 21

Figure 22

Figure 23

Figure 24

Figure 25

Figure 26

Figure 27

Embodiments for Carrying Out the Invention

[0011] Hereinafter, embodiments of a learning system, an estimation system, a learning method, and a computer program will be described with reference to the drawings.

[0012] <First Embodiment> The learning system according to the first embodiment will be described with reference to FIGS. 1 to 4.

[0013] (Hardware Configuration) First, with reference to FIG. 1, the hardware configuration of the learning system 10 according to the first embodiment will be described. FIG. 1 is a block diagram showing the hardware configuration of the learning system according to the first embodiment.

[0014] As shown in FIG. 1, the learning system 10 according to the first embodiment includes a processor 11, a RAM (Random Access Memory) 12, a ROM (Read Only Memory) 13, and a storage device 14. The learning system 10 may further include an input device 15 and an output device 16. The processor 11, the RAM 12, the ROM 13, the storage device 14, the input device 15, and the output device 16 are connected via a data bus 17.

[0015] Processor 11 reads a computer program. For example, processor 11 is configured to read a computer program stored in at least one of RAM 12, ROM 13, and storage device 14. Alternatively, processor 11 may read a computer program stored in a computer-readable recording medium using a recording medium reading device (not shown). Processor 11 may obtain (i.e., read) a computer program from a device (not shown) disposed outside learning system 10 via a network interface. By executing the read computer program, processor 11 controls RAM 12, storage device 14, input device 15, and output device 16. In particular, in this embodiment, when the computer program read by processor 11 is executed, function blocks for executing processing related to machine learning are realized in processor 11. Also, as processor 11, one of a CPU (Central Processing Unit), GPU (Graphics Processing Unit), FPGA (field-programmable gate array), DSP (Demand-Side Platform), and ASIC (Application Specific Integrated Circuit) may be used, or a plurality may be used in parallel.

[0016] RAM 12 temporarily stores the computer program executed by processor 11. RAM 12 temporarily stores data temporarily used by processor 11 when processor 11 is executing a computer program. RAM 12 may be, for example, D-RAM (Dynamic RAM).

[0017] ROM 13 stores the computer program executed by processor 11. ROM 13 may also store other fixed data. ROM 13 may be, for example, P-ROM (Programmable ROM).

[0018] The storage device 14 stores data that the learning system 10 stores in the long term. The storage device 14 may operate as a temporary storage device of the processor 11. The storage device 14 may include, for example, at least one of a hard disk device, a magneto-optical disk device, an SSD (Solid State Drive), and a disk array device.

[0019] The input device 15 is a device that receives an input instruction from the user of the learning system 10. The input device 15 may include, for example, at least one of a keyboard, a mouse, and a touch panel.

[0020] The output device 16 is a device that outputs information regarding the learning system 10 to the outside. For example, the output device 16 may be a display device (e.g., a display) capable of displaying information regarding the learning system 10.

[0021] (Functional configuration) Next, with reference to FIG. 2, the functional configuration of the learning system 10 according to the first embodiment will be described. FIG. 2 is a block diagram showing the functional configuration of the learning system according to the first embodiment.

[0022] As shown in FIG. 2, the learning system 10 according to the first embodiment includes a composite image generation unit 110, an iris information estimation unit 120, and a learning unit 130 as processing blocks for realizing its functions. Each of the composite image generation unit 110, the iris information estimation unit 120, and the learning unit 130 may be realized, for example, in the above-described processor 11 (see FIG. 1).

[0023] The composite image generation unit 110 is configured to generate a composite image by synthesizing information regarding the reflection component of light not derived from the original image onto the original image including a living body. Here, the "original image" is the image before synthesizing the reflection component. Also, the "information regarding the reflection component" is information related to the component generated by light such as natural light or illumination light being reflected by an object with a high reflectivity such as a mirror, glass, or glasses. The "information regarding the reflection component" may be information indicating the reflection component itself or information for simulating the reflection component. Hereinafter, the "information regarding the reflection component" will be appropriately referred to as the "reflection component". The composite image generation unit 110 acquires the original image and the reflection component respectively, and executes a process of synthesizing the reflection component onto the original image. Thus, the composite image is an image in which the reflection component exists in at least a part of the region of the original image. Note that the synthesis process may be a simple overlapping process or a more complex synthesis process. Since existing techniques can be appropriately adopted for the synthesis process, detailed description thereof is omitted here. The composite image generated by the composite image generation unit 110 is configured to be output to the iris information estimation unit 120.

[0024] The iris information estimation unit 120 is configured to estimate information regarding a specific part of the living body based on the composite image (i.e., the image with the reflection component synthesized) generated by the composite image generation unit 110. The specific part of the living body may be, for example, a part used for biometric authentication. The specific part of the living body is not particularly limited, but in this embodiment, as an example, information regarding the iris is estimated, and hereinafter, the description will proceed assuming that the specific part of the living body is the iris. The information regarding the specific part of the living body may be, for example, information indicating the position of the specific part (coordinate information, etc.) or information indicating the feature amount of the specific part. The iris information estimation unit 120 may estimate only one type of information regarding the specific part of the living body or may estimate a plurality of types. Note that since existing techniques can be appropriately adopted for the method of estimating information regarding the specific part of the living body from the image, detailed description thereof is omitted here. The information regarding the specific part of the living body estimated by the iris information estimation unit 120 is configured to be output to the learning unit 130.

[0025] The learning unit 130 performs learning of the iris information estimation unit 120 based on the information regarding a specific part of the living body estimated by the iris information estimation unit 120 and the correct information indicating the correct answer of the information regarding the specific part in the original image. Specifically, the learning unit 130 optimizes the parameters based on the information regarding the iris estimated by the iris information estimation unit 120 and the correct information regarding the iris in the original image so that the iris information estimation unit 120 can estimate information regarding the iris with higher accuracy. Here, the "correct information" refers to the information regarding the specific part actually included in the original image, and is preset as the information to be estimated by the iris information estimation unit 120 from the synthesized image. Specific examples of the correct information include the central position of the iris, the radius of the iris circle, the feature points on the iris circle, the central position of the pupil, the radius of the pupil, the feature points on the pupil circle, the feature points of the eyelid, eye opening and closing, the length of the eyelashes, the number of eyelashes, the density of eyelashes, the presence or absence of false eyelashes, double eyelids or not, the presence or absence of makeup, the presence or absence of glasses, the presence or absence of contact lenses, the presence or absence of eye abnormalities, the information of whether it is a living body or not, the color of the iris, the type of iris pattern, the information for identifying an individual, etc. The correct information is input as training data paired with the original image. The correct information may be created by human work or may be information estimated from the original image by a method different from the iris information estimation unit 120. The learning unit 130 typically executes learning of the iris information estimation unit 120 using a plurality of synthesized images and their correct information. Note that since existing techniques can be appropriately adopted for the specific method of learning by the learning unit 130, detailed description thereof is omitted here.

[0026] (Synthesis of reflection components) Next, with reference to FIG. 3, the synthesis of the reflection components by the above-described synthesis image generation unit 110 will be specifically described. FIG. 3 is a conceptual diagram showing an example of the synthesis process of the original image and the reflection components.

[0027] The composite image generation unit 110 receives an original image (iris image) and a reflection component as shown in FIG. 3. Here, the reflection component may be, for example, an assumed reflection component when it is assumed that the living body wears glasses and light is reflected by the glasses. As shown in the figure, the reflection component may be a component having a relatively high pixel value (here, an image having three rectangular regions with a color close to white as the reflection component is input. The shape of the region of the reflection component is not limited to a rectangular region and may be any figure).

[0028] The composite image generation unit 110 synthesizes the input original image and reflection component to generate a composite image. As shown in FIG. 3, the composite image may be an image in which the reflection component (i.e., the white rectangular region) is superimposed on the original image. Here, one composite image is generated from one original image, but a plurality of composite images may be generated from one original image. In that case, different reflection components may be synthesized on one original image, or the same reflection component may be synthesized in different manners. The configuration of synthesizing the same reflection component in different manners will be described in detail in other embodiments described later.

[0029] (Flow of operations) Next, with reference to FIG. 4, the flow of operations of the learning system 10 according to the first embodiment will be described. FIG. 4 is a flowchart showing the flow of operations of the learning system according to the first embodiment.

[0030] As shown in FIG. 4, when the learning system 10 according to the first embodiment operates, first, the composite image generation unit 110 acquires a set of original images as an image dataset (step S101). The composite image generation unit 110 may acquire, for example, an image dataset stored in a storage or the like. After acquiring the image dataset, it enters the learning loop. In the learning loop, first, training data composed of pairs of original images and correct answers are randomly acquired from the acquired image dataset in a predetermined batch size number (step S150).

[0031] Subsequently, the composite image generation unit 110 synthesizes the reflection component onto the original image included in the acquired batch (step S102). The composite image generation unit 110 outputs the generated composite image to the iris information estimation unit 120. The composite image generation unit 110 may generate a plurality of composite images and output the plurality of composite images to the iris information estimation unit 120 collectively.

[0032] Subsequently, the iris information estimation unit 120 estimates information regarding the iris based on the composite image generated by the composite image generation unit 110 (step S103). Then, the learning unit 130 performs learning processing of the iris information estimation unit 120 based on the information regarding the iris estimated by the iris information estimation unit 120 (hereinafter, appropriately referred to as "estimated information") and the correct information in the original image (step S104).

[0033] Subsequently, the learning unit 130 determines whether or not all learning has been completed (step S105). The learning unit 130 may determine that learning has been completed, for example, when a predetermined number of iterations has been reached. Alternatively, the learning unit 130 may determine that learning has been completed when a predetermined estimation accuracy has been reached, or when the change in the loss calculated based on the estimated information and the correct information has become constant.

[0034] When it is determined that learning has been completed (step S105: YES), the series of processes ends. On the other hand, when it is determined that learning has not been completed (step S105: NO), the process may be started again from step S150 (that is, the part that acquires training data of the batch size). In the processing of step S102 for the second and subsequent times, a composite image may be generated by synthesizing the reflection component onto a different original image. Also, different composite images may be generated for each loop, such as synthesizing the reflection component in a different manner for the same original image.

[0035] (Technical Effect) Next, the technical effect obtained by the learning system 10 according to the first embodiment will be described.

[0036] As described with reference to FIGS. 1 to 4, in the learning system 10 according to the first embodiment, the iris information estimation unit 120 is trained using a composite image obtained by synthesizing a reflection component with the original image. By training the iris information estimation unit 120 in this way, even when a reflection component is included in an image including a living body (for example, when the reflection component overlaps a specific part of the living body), it is possible to accurately estimate information regarding the specific part of the living body. Further, by generating a plurality of composite images from a single original image, the images used for training can be expanded, so that a model more robust to reflection components can be trained.

[0037] <Modification Example> The learning system 10 according to a modification example of the first embodiment will be described with reference to FIGS. 5 and 6. Note that the modification example described below differs only in some configurations and operations from the first embodiment, and the other parts may be the same as those of the first embodiment (see FIGS. 1 to 4). Therefore, below, the parts different from the first embodiment already described will be described in detail, and the description of the other overlapping parts will be omitted as appropriate.

[0038] (Configuration of the Modification Example) First, with reference to FIG. 5, the functional configuration of the learning system 10 according to the modification example of the first embodiment will be described. FIG. 5 is a block diagram showing the functional configuration of the learning system according to the modification example of the first embodiment. In FIG. 5, the same reference numerals are given to the elements similar to those shown in FIG. 2.

[0039] As shown in FIG. 5, the learning system 10 according to the modification example of the first embodiment includes, as processing blocks for realizing its functions, a composite image generation unit 110, an iris information estimation unit 120, and a learning unit 130. In particular, in the learning system 10 according to the modification example, the learning unit 130 includes a loss function calculation unit 131, a gradient calculation unit 132, and a parameter update unit 133.

[0040] The loss function calculation unit 131 is configured to be able to calculate a loss function based on the error between the estimated information estimated by the iris information estimation unit 120 and the correct information in the original image. Regarding the specific calculation method of the loss function, since existing technologies can be appropriately adopted, detailed description here is omitted.

[0041] The gradient calculation unit 132 is configured to be able to calculate a gradient using the loss function calculated by the loss function calculation unit 131. Regarding the specific calculation method of the gradient, since existing technologies can be appropriately adopted, detailed description here is omitted.

[0042] The parameter update unit 133 is configured to be able to update the parameters in the iris information estimation unit 120 (that is, the parameters for estimating iris information) based on the gradient calculated by the gradient calculation unit 132. The parameter update unit 133 optimizes the parameters so that the estimated information is estimated as information closer to the correct information by updating the parameters so that the loss calculated by the loss function becomes smaller.

[0043] (Operation of the modification example) Next, with reference to FIG. 6, the operation flow of the learning system 10 according to the modification example of the first embodiment will be described. FIG. 6 is a flowchart showing the operation flow of the learning system according to the modification example of the first embodiment. In FIG. 6, the same reference numerals are given to the same processes as those shown in FIG. 4.

[0044] As shown in FIG. 6, when the learning system 10 according to the modification example of the first embodiment operates, first, the composite image generation unit 110 acquires an image dataset composed of the original image (step S101). Then, training data composed of a pair of the original image and the correct answer is acquired from the acquired image dataset in a predetermined number of batch sizes (step S150). Thereafter, the composite image generation unit 110 synthesizes a reflection component with the acquired original image (step S102).

[0045] Subsequently, the iris information estimation unit 120 estimates information regarding the iris based on the composite image generated by the composite image generation unit 110 (step S103). The estimated information estimated by the iris information estimation unit 120 is output to the loss function calculation unit 131 in the learning unit 130.

[0046] Subsequently, the loss function calculation unit 131 calculates a loss function based on the estimated information input from the iris information estimation unit 120 and the correct information acquired as training data (step S111). Then, the gradient calculation unit 132 calculates a gradient using the loss function (step S112). Thereafter, the parameter update unit 133 updates the parameters of the iris information estimation unit 120 based on the calculated gradient (step S113).

[0047] Subsequently, the learning unit 130 determines whether or not all learning has been completed (step S105). If it is determined that the learning has been completed (step S105: YES), the series of processes ends. On the other hand, if it is determined that the learning has not been completed (step S105: NO), the process may be started again from step S150 (i.e., the part where training data of the batch size is acquired).

[0048] (Effect of the modification example) Next, the technical effects obtained by the learning system 10 according to the modification example of the first embodiment will be described.

[0049] As described with reference to FIGS. 5 and 6, in the learning system 10 according to the first embodiment, the parameters in the iris information estimation unit 120 are optimized based on the gradient calculated using the loss function. Even when the iris information estimation unit 120 is learned in this way, as in the learning system 10 according to the first embodiment described above, even when the reflected component is included in the image including the living body, it is possible to accurately estimate the information regarding the specific part of the living body. In addition, since a plurality of composite images can be generated from one original image to expand the images used for learning, a model more robust to the reflected component can be learned.

[0050] <Second Embodiment> The learning system 10 according to the second embodiment will be described with reference to FIGS. 7 to 9. Note that the second embodiment is only different from the above-described first embodiment in some configurations and operations, and the other parts may be the same as those of the first embodiment (see FIGS. 1 and 2). Therefore, hereinafter, the description of the parts overlapping with the first embodiment already described will be omitted as appropriate.

[0051] (Functional configuration) Next, with reference to FIG. 7, the functional configuration of the learning system 10 according to the second embodiment will be described. FIG. 7 is a block diagram showing the functional configuration of the learning system according to the second embodiment. In FIG. 7, the same reference numerals are given to the elements similar to those shown in FIG. 2.

[0052] As shown in FIG. 7, the learning system 10 according to the second embodiment includes, as processing blocks for realizing its functions, a composite image generation unit 110, an iris information estimation unit 120, a learning unit 130, and a reflection component extraction unit 140. That is, the learning system 10 according to the second embodiment is configured to further include a reflection component extraction unit 140 in addition to the configuration of the first embodiment (see FIG. 2). The reflection component extraction unit 140 may be realized, for example, in the above-described processor 11 (see FIG. 1).

[0053] The reflection component extraction unit 140 is configured to be able to extract a reflection component from an image including a reflection component (hereinafter, appropriately referred to as a "reflection image"). The reflection image here only needs to include a reflection component to some extent, and the imaging object can be anything. For example, the reflection image may be an image of the same living body (i.e., the same person) as the living body included in the original image, or an image of another living body. Alternatively, the reflection image may be an image that does not include a living body. The specific method for extracting the reflection component by the reflection component extraction unit 140 will be described in detail in other embodiments described later. The reflection component extracted by the reflection component extraction unit 140 is configured to be output to the composite image generation unit 110.

[0054] (Synthesis of Reflection Components) Next, with reference to FIG. 8, the extraction of the reflection component by the above-described reflection component extraction unit 140 will be specifically described. FIG. 8 is a conceptual diagram showing an example of the synthesis process of the reflection component extracted from the original image and the reflection image.

[0055] As shown in FIG. 8, the reflection component extraction unit 140 extracts only the reflection component from the reflection image including the reflection component. The reflection image may be, for example, an image of a person wearing glasses as shown in the figure. Such a reflection image includes a reflection component caused by light being reflected by the glasses. Alternatively, the reflection image may be an image of only the glasses (i.e., an image not including a person). Also in this case, the reflection image includes a reflection component caused by light being reflected by the glasses. Note that the reflection image may not include an object that directly reflects light such as glasses. For example, it may be an image including a light source such as illumination instead of a reflection object, or an image including light reflected by a light source outside the visual field within the lens.

[0056] (Flow of Operations) Next, with reference to FIG. 9, the flow of operations of the learning system 10 according to the second embodiment will be described. FIG. 9 is a flowchart showing the flow of operations of the learning system according to the second embodiment. In FIG. 9, the same reference numerals are assigned to the same processes as those shown in FIG. 4.

[0057] As shown in FIG. 9, when the learning system 10 according to the second embodiment operates, first, the composite image generation unit 110 acquires an image data set of the original image (step S101).

[0058] Subsequently, the reflection component extraction unit 140 acquires a reflection image data set (step S201). The reflection component extraction unit 140 may acquire, for example, an image data set including the reflection component stored in a storage or the like.

[0059] Subsequently, the reflection component extraction unit 140 extracts the reflection components from each of the acquired reflection image datasets (step S202). The reflection component extraction unit 140 outputs the extracted reflection components to the composite image generation unit 110. Note that the processes of step S201 and step S202 described above may be performed in parallel with the process of step S101 (i.e., the process in which the composite image generation unit 110 acquires the image dataset of the original image), or may be performed before that.

[0060] Subsequently, the composite image generation unit 110 acquires training data composed of pairs of the original image and the correct answer from among the acquired image datasets, in a number corresponding to a predetermined batch size (step S150). Then, the composite image generation unit 110 synthesizes the reflection components extracted by the reflection component extraction unit 140 with the acquired original image (step S102). The composite image generation unit 110 outputs the generated composite image to the iris information estimation unit 120.

[0061] Subsequently, the iris information estimation unit 120 estimates information regarding the iris based on the composite image generated by the composite image generation unit 110 (step S103). Then, the learning unit 130 performs the learning process of the iris information estimation unit 120 based on the estimated information estimated by the iris information estimation unit 120 and the correct answer information in the original image (step S104).

[0062] Subsequently, the learning unit 130 determines whether or not all learning has been completed (step S105). If it is determined that the learning has been completed (step S105: YES), the series of processes ends. On the other hand, if it is determined that the learning has not been completed (step S105: NO), the process may simply be restarted from step S150. Note that when a plurality of types of reflection components are extracted in step S202 (i.e., when a plurality of reflection components are extracted), in the process of step S102 for the second and subsequent times, different reflection components may be synthesized to generate a composite image.

[0063] (Technical Effect) Next, the technical effect obtained by the learning system 10 according to the second embodiment will be described.

[0064] As described with reference to FIGS. 7 to 9, in the learning system 10 according to the second embodiment, a reflection component is extracted from a reflection image, and a composite image is generated by synthesizing the reflection components. In this way, the reflection components used for synthesis with the original image can be obtained by a predetermined extraction process. Further, since the reflection components that may occur when actually imaging an image are extracted, the data at the time of learning can be augmented using the extracted reflection components, and a model that is more robust to environmental changes can be learned.

[0065] <Modification Example of the Second Embodiment> The learning system 10 according to the modification example of the second embodiment will be described with reference to FIG. 10. Note that the modification example described below differs only in some operations from the second embodiment, and the other parts may be the same as those of the second embodiment. For this reason, below, the parts different from the second embodiment already described will be described in detail, and the description of the other overlapping parts will be omitted as appropriate.

[0066] (Operation of the Modification Example) First, with reference to FIG. 10, the flow of the operation of the learning system 10 according to the modification example of the second embodiment will be described. FIG. 10 is a flowchart showing the flow of the operation of the learning system according to the modification example of the second embodiment. Note that in FIG. 10, the same reference numerals are given to the processes similar to those shown in FIG. 9.

[0067] As shown in FIG. 10, when the learning system 10 according to the modification example of the second embodiment operates, first, the composite image generation unit 110 acquires an image data set of the original image (step S101).

[0068] Subsequently, the reflection component extraction unit 140 acquires a reflection image data set (step S201). Then, the reflection component extraction unit 140 acquires a predetermined number of reflection images from the reflection image data set in terms of the batch size (step S151).

[0069] Subsequently, the reflection component extraction unit 140 extracts the reflection component from the acquired reflection image (step S202). The reflection component extraction unit 140 outputs the extracted reflection component to the composite image generation unit 110.

[0070] Subsequently, the composite image generation unit 110 acquires training data composed of pairs of the original image and the correct answer from the acquired image dataset by a predetermined number of batch sizes (step S150). Then, the composite image generation unit 110 composites the reflection component extracted by the reflection component extraction unit 140 with the acquired original image (step S102). The composite image generation unit 110 outputs the generated composite image to the iris information estimation unit 120.

[0071] Subsequently, the iris information estimation unit 120 estimates information regarding the iris based on the composite image generated by the composite image generation unit 110 (step S103). Then, the learning unit 130 performs learning processing of the iris information estimation unit 120 based on the estimated information estimated by the iris information estimation unit 120 and the correct answer information in the original image (step S104).

[0072] Subsequently, the learning unit 130 determines whether all learning has been completed (step S105). If it is determined that the learning has been completed (step S105: YES), the series of processes ends. On the other hand, if it is determined that the learning has not been completed (step S105: NO), the process may be started again from step S201.

[0073] (Effect of the modification example) Next, the technical effects obtained by the learning system 10 according to the modification example of the second embodiment will be described.

[0074] As described with reference to FIG. 10, in the learning system 10 according to the modified example of the second embodiment, the reflection component is extracted each time the process loops, and the composite image is generated by synthesizing the extracted reflection components. Also in such a case, similar to the second embodiment already described, the reflection component used for synthesis with the original image can be obtained by a predetermined extraction process. Further, since the reflection component that may occur when actually imaging an image is extracted, the data at the time of learning can be augmented using the extracted reflection component, and a model more robust to environmental changes can be learned.

[0075] <Third Embodiment> The learning system 10 according to the third embodiment will be described with reference to FIGS. 11 and 12. Note that the third embodiment is different only in some configurations and operations as compared with the first and second embodiments described above, and the other parts may be the same as those of the first and second embodiments. For this reason, hereinafter, the description of the parts overlapping with the embodiments already described will be omitted as appropriate.

[0076] (Functional Configuration) First, the functional configuration of the learning system 10 according to the third embodiment will be described with reference to FIG. 11. FIG. 11 is a block diagram showing the functional configuration of the learning system according to the third embodiment. In FIG. 11, the same reference numerals are given to the elements similar to those shown in FIG. 7.

[0077] As shown in FIG. 11, the learning system 10 according to the third embodiment includes, as processing blocks for realizing its functions, a composite image generation unit 110, an iris information estimation unit 120, a learning unit 130, and a reflection component extraction unit 140. In particular, in the learning system 10 according to the third embodiment, the reflection component extraction unit 140 includes a pixel value determination unit 141.

[0078] The pixel value determination unit 141 is configured to determine the pixel value of each pixel of the reflected image and extract a pixel (region) whose pixel value is equal to or greater than a predetermined threshold as a reflection component. Here, the "predetermined threshold" is a threshold for determining a pixel including a reflection component, and is set as, for example, a hyperparameter of learning. The predetermined threshold may be set as a value between, for example, 200 and 255 (that is, a value that can discriminate pixels with relatively high pixel values).

[0079] The predetermined threshold may be a variable value. Also, a plurality of predetermined thresholds may be prepared and selected and used according to the situation. In this case, the predetermined threshold may be a value that can be changed and selected by the system administrator or the like. Further, the predetermined threshold may be set as a value that is changed and selected according to the input reflected image. For example, the predetermined threshold may be a value linked to the average pixel value of the reflected image or the like. When the reflected image is a relatively dark image, the predetermined threshold may be a low value, and when the reflected image is a relatively bright image, the predetermined threshold may be a high value.

[0080] (Flow of operations) Next, with reference to FIG. 12, the flow of operations of the learning system 10 according to the third embodiment will be described. FIG. 12 is a flowchart showing the flow of operations of the learning system according to the third embodiment. In FIG. 12, the same reference numerals are given to the same processes as those shown in FIG. 9.

[0081] As shown in FIG. 12, when the learning system 10 according to the third embodiment operates, first, the composite image generation unit 110 acquires an image data set of the original image (step S101).

[0082] Subsequently, the reflection component extraction unit 140 acquires a reflected image (step S201). Then, the pixel value determination unit 141 determines the pixel value of the acquired reflected image (step S301). And the pixel value determination unit 141 extracts a region including pixels whose pixel value is equal to or greater than a predetermined value as a reflection component (step S302). Note that there may be a plurality of reflected images. In that case, reflection components may be generated from each reflected image respectively.

[0083] Subsequently, the composite image generation unit 110 acquires training data composed of pairs of the original image and the correct answer from the acquired image dataset by a predetermined number of batch sizes (step S150). Then, the composite image generation unit 110 synthesizes the reflection component extracted by the reflection component extraction unit 140 (that is, the component corresponding to the pixel determined to have a pixel value equal to or greater than a predetermined value by the pixel value determination unit 141) with the acquired original image (step S102). The composite image generation unit 110 outputs the generated composite image to the iris information estimation unit 120. When there are a plurality of reflection components, they may be randomly selected from the plurality of reflection images and synthesized with the original image.

[0084] Subsequently, the iris information estimation unit 120 estimates information regarding the iris based on the composite image generated by the composite image generation unit 110 (step S103). Then, the learning unit 130 performs learning processing of the iris information estimation unit 120 based on the estimated information estimated by the iris information estimation unit 120 and the correct answer information in the original image (step S104).

[0085] Subsequently, the learning unit 130 determines whether all learning has been completed (step S105). If it is determined that the learning has been completed (step S105: YES), the series of processes ends. On the other hand, if it is determined that the learning has not been completed (step S105: NO), the process may be started again from step S150.

[0086] (Technical Effect) Next, the technical effect obtained by the learning system 10 according to the third embodiment will be described.

[0087] As described with reference to FIGS. 11 and 12, in the learning system 10 according to the third embodiment, a region where the pixel value in the reflection image is a predetermined threshold value is extracted as a reflection component. By doing so, since the reflection component extracted from the reflection image is combined with the original image to obtain more difficult training data, a model robust to the reflection component can be learned. In the above-described embodiment, an example of extracting by comparing the pixel value with the threshold value has been given. However, when it is possible to determine whether it is a reflection component based on a parameter other than the pixel value, a threshold value for the parameter may be set to extract the reflection component.

[0088] <Modification Example of the Third Embodiment> The learning system 10 according to the modification example of the third embodiment will be described with reference to FIG. 13. Note that the modification example described below is different only in some operations from the third embodiment, and other parts may be the same as those of the third embodiment. For this reason, below, the parts different from the third embodiment already described will be described in detail, and the description of other overlapping parts will be omitted as appropriate.

[0089] (Operation of the Modification Example) First, with reference to FIG. 13, the flow of the operation of the learning system 10 according to the modification example of the third embodiment will be described. FIG. 13 is a flowchart showing the flow of the operation of the learning system according to the modification example of the third embodiment. In FIG. 13, the same reference numerals are given to the same processes as those shown in FIG. 12.

[0090] As shown in FIG. 13, when the learning system 10 according to the third embodiment operates, first, the composite image generation unit 110 acquires an image data set of the original image (step S101).

[0091] Subsequently, the reflection component extraction unit 140 acquires a data set of the reflection image (step S201). Then, the reflection component extraction unit 140 acquires a predetermined number of reflection images from the reflection image data set in a batch size (step S151).

[0092] Subsequently, the pixel value determination unit 141 determines the pixel values of the acquired reflected image (step S301). Then, the pixel value determination unit 141 extracts, as a reflection component, a region including pixels whose pixel values are equal to or greater than a predetermined value (step S302).

[0093] Subsequently, the composite image generation unit 110 acquires, from the acquired image data set, training data composed of pairs of the original image and the correct answer, in a number corresponding to a predetermined batch size (step S150). Then, the composite image generation unit 110 composites the reflection component extracted by the reflection component extraction unit 140 (that is, the component corresponding to the pixels determined to have pixel values equal to or greater than the predetermined value by the pixel value determination unit 141) with the acquired original image (step S102). The composite image generation unit 110 outputs the generated composite image to the iris information estimation unit 120.

[0094] Subsequently, the iris information estimation unit 120 estimates information regarding the iris based on the composite image generated by the composite image generation unit 110 (step S103). Then, the learning unit 130 performs learning processing of the iris information estimation unit 120 based on the estimated information estimated by the iris information estimation unit 120 and the correct answer information in the original image (step S104).

[0095] Subsequently, the learning unit 130 determines whether all learning has been completed (step S105). If it is determined that the learning has been completed (step S105: YES), the series of processes ends. On the other hand, if it is determined that the learning has not been completed (step S105: NO), the processing may simply be restarted from step S201.

[0096] (Effect of the modification) Next, the technical effects obtained by the learning system 10 according to the modification of the third embodiment will be described.

[0097] As described with reference to FIG. 13, in the learning system 10 according to the modification of the third embodiment, a reflection component is acquired each time the process loops, and a region in the reflected image where the pixel value is a predetermined threshold is extracted as the reflection component. By doing so, different reflection components are acquired for each batch, so the difficulty level of the training data increases, and a model that is more robust to the reflection component can be learned.

[0098] <Fourth Embodiment> The learning system 10 according to the fourth embodiment will be described with reference to FIGS. 14 to 16. Note that the fourth embodiment may be the same as the first to third embodiments described above except that some configurations and operations are different, and other parts may be the same as the first to third embodiments. For this reason, hereinafter, descriptions of parts overlapping with the embodiments already described will be omitted as appropriate.

[0099] (Functional Configuration) First, with reference to FIG. 14, the functional configuration of the learning system 10 according to the fourth embodiment will be described. FIG. 14 is a block diagram showing the functional configuration of the learning system according to the fourth embodiment. In FIG. 14, the same reference numerals are given to the same elements as those shown in FIG. 2.

[0100] As shown in FIG. 14, the learning system 10 according to the fourth embodiment includes a composite image generation unit 110, an iris information estimation unit 120, and a learning unit 130 as processing blocks for realizing its functions. In particular, in the learning system 10 according to the third embodiment, the composite image generation unit 110 includes a perturbation processing unit 111.

[0101] The perturbation processing unit 111 is configured to be able to execute various processes for applying perturbations to the reflection component to be synthesized with the original image. Specifically, the perturbation processing unit 111 is configured to be able to execute at least one of the processes of "moving", "rotating", "enlarging", and "shrinking" on the reflection component. The processes executed by the perturbation processing unit 111 will be specifically described below.

[0102] (Examples of Perturbation Processing) Next, with reference to FIG. 15, a specific example of the perturbation process by the above-described perturbation processing unit 111 will be described. FIG. 15 is a conceptual diagram showing an example of the perturbation process for the reflection component.

[0103] As shown in FIG. 15, the perturbation processing unit 111 may be configured to be capable of executing a movement process of the reflection component. In the movement process, the position of the reflection component is moved in the vertical, horizontal, or diagonal direction. The amount of movement and the direction of movement in the movement process may be set in advance, may be appropriately changed by the user or the like, or may be randomly set according to a random number. When the movement process is executed multiple times, the amount of movement and the direction of movement may be the same each time, or may be changed each time.

[0104] The perturbation processing unit 111 may be configured to be capable of executing a rotation process of the reflection component. In the rotation process, the reflection component is rotated clockwise or counterclockwise. Alternatively, the rotation process may perform a three-dimensional rotation including the depth direction of the image. Also, the axis of rotation may be the center of the image, or may be other coordinates. The axis of rotation may be an inner coordinate or an outer coordinate of the image. The amount of rotation, the direction of rotation, and the coordinates of the axis of rotation in the rotation process may be set in advance, may be appropriately changed by the user or the like, or may be randomly set according to a random number. When the rotation process is executed multiple times, the amount of rotation, the direction of rotation, and the coordinates of the axis of rotation may be the same each time, or may be changed each time.

[0105] The perturbation processing unit 111 may be configured to be capable of executing an enlargement process of the reflection component. In the enlargement process, the size of the reflection component is changed to be larger. The amount of enlargement and the enlargement ratio in the enlargement process may be set in advance, may be appropriately changed by the user or the like, or may be randomly set according to a random number. When the enlargement process is executed multiple times, the amount of enlargement and the enlargement ratio may be the same each time, or may be changed each time.

[0106] The perturbation processing unit 111 may be configured to be capable of executing reduction processing on the reflection component. In the reduction processing, the size of the reflection component is changed to be smaller. The reduction amount and reduction rate in the reduction processing may be set in advance, may be appropriately changed by the user or the like, or may be randomly set according to a random number. When the reduction processing is executed multiple times, the reduction amount and reduction rate may be the same each time, or may be changed each time.

[0107] The above-described movement processing, rotation processing, enlargement processing, and reduction processing may be executed in combination. For example, by simultaneously executing the movement processing and the rotation processing, both the position and the angle of the reflection component may be changed. Alternatively, by simultaneously executing the rotation processing and the enlargement processing, both the angle and the size of the reflection component may be changed. Alternatively, by simultaneously executing the movement processing and the reduction processing, both the position and the size of the reflection component may be changed. Note that the combination of each process may be appropriately changed. For example, it may be determined which operations are to be combined and executed according to the user's operation, or it may be changed randomly.

[0108] Also, when the perturbation processing is executed in multiple parts, different processes may be executed sequentially. For example, the movement processing may be executed for the first time, the rotation processing may be executed for the second time, the enlargement processing may be executed for the third time, and the reduction processing may be executed for the fourth time. Note that regarding which process is to be executed in what order, it may be appropriately changed. For example, the order of the processes may be determined according to the user's operation, or the order may be determined randomly.

[0109] When the reflection component is divided into a plurality of regions, separate perturbation processing may be executed for each region. For example, as shown in FIG. 15, when the reflection component is divided into three rectangular regions, movement processing may be executed for the first rectangular region, rotation processing may be executed for the second rectangular region, and enlargement processing may be executed for the third rectangular region. Specifically, each region may be arranged in a different channel, and a channel for executing perturbation processing and a channel for not executing it may be selected by random numbers. Note that which processing is executed for which region may be changed as appropriate. For example, according to the operation of the user, which processing is executed for which region may be determined, or it may be determined randomly.

[0110] The perturbation processing unit 111 may always execute perturbation processing, or may omit perturbation processing as appropriate. For example, for the synthetic image generated first, perturbation processing may not be executed, and for the synthetic image generated second, perturbation processing may be executed. When generating a synthetic image, the presence or absence of execution may be determined using random numbers. For example, using a uniform random number in the range from 1 to 0, when the obtained random number value is 0.5 or more, perturbation processing may be executed, and in other cases, perturbation processing may not be executed. Alternatively, when the reflection component is divided into a plurality of regions, there may be a region where perturbation processing is executed and a region where perturbation processing is not executed. For example, as shown in FIG. 15, when the reflection component is divided into three rectangular regions, perturbation processing may not be executed for the first rectangular region, and perturbation processing may be executed only for the second and third rectangular regions. Alternatively, perturbation processing may be executed for the first and second rectangular regions, and perturbation processing may not be executed only for the third rectangular region. Note that whether to execute perturbation processing or not may be determined randomly.

[0111] (Flow of operation) Next, with reference to FIG. 16, the flow of operation of the learning system 10 according to the fourth embodiment will be described. FIG. 16 is a flowchart showing the flow of operation of the learning system according to the fourth embodiment. In FIG. 16, the same reference numerals are given to the same processes as those shown in FIG. 4.

[0112] As shown in FIG. 16, when the learning system 10 according to the fourth embodiment operates, first, the composite image generation unit 110 acquires an image data set of the original image (step S101). Then, training data composed of a pair of the original image and the correct answer is acquired from the acquired image data set by a predetermined number of batch sizes (step S150).

[0113] Subsequently, the perturbation processing unit 111 performs perturbation processing on the reflection component to apply perturbation (step S401). Then, the composite image generation unit 110 synthesizes the reflection component to which perturbation has been applied with the acquired original image (step S102). The composite image generation unit 110 outputs the generated composite image to the iris information estimation unit 120.

[0114] Subsequently, the iris information estimation unit 120 estimates information regarding the iris based on the composite image generated by the composite image generation unit 110 (step S103). Then, the learning unit 130 performs learning processing of the iris information estimation unit 120 based on the estimated information regarding the iris estimated by the iris information estimation unit 120 and the correct answer information in the original image (step S104).

[0115] Subsequently, the learning unit 130 determines whether all learning has been completed (step S105). If it is determined that the learning has been completed (step S105: YES), the series of processes ends. On the other hand, if it is determined that the learning has not been completed (step S105: NO), the process may be started again from step S150. As already described, in the processing of step S401 for the second and subsequent times, perturbation processing different from the first time may be executed.

[0116] (Technical Effect) Next, the technical effect obtained by the learning system 10 according to the fourth embodiment will be described.

[0117] As described with reference to FIGS. 14 to 16, in the learning system 10 according to the fourth embodiment, perturbation processing is executed on the reflection component before synthesis. By doing so, it becomes possible to synthesize one reflection component into the original image in different modes. Therefore, a plurality of synthesized images can be generated from one original image and one reflection component, and it is possible to efficiently expand the images used for learning.

[0118] <Fifth Embodiment> The learning system 10 according to the fifth embodiment will be described with reference to FIGS. 17 to 19. Note that the fifth embodiment may be the same as the first to fourth embodiments described above except that some configurations and operations are different, and the other parts may be the same as those of the first to fourth embodiments. For this reason, hereinafter, descriptions of parts overlapping with the embodiments already described will be omitted as appropriate. (Functional Configuration) First, with reference to FIG. 17, the functional configuration of the learning system 10 according to the fifth embodiment will be described. FIG. 17 is a block diagram showing the functional configuration of the learning system according to the fifth embodiment. In FIG. 17, the same reference numerals are given to the same elements as those shown in FIG. 2.

[0119] As shown in FIG. 17, the learning system 10 according to the fifth embodiment includes a synthesized image generation unit 110, an iris information estimation unit 120, and a learning unit 130 as processing blocks for realizing its functions. In particular, in the learning system 10 according to the fifth embodiment, the synthesized image generation unit 110 includes an overlap determination unit 112.

[0120] The overlap determination unit 112 is configured to be able to determine whether or not a specific part of the living body and the reflection component overlap at least partially when the reflection component is synthesized into the original image. Whether or not the specific part of the living body and the reflection component overlap can be made possible, for example, by acquiring the position information of the specific part of the living body in advance. The position information of the specific part may be, for example, the correct answer information input to the learning unit 130 or the information estimated from the original image.

[0121] In addition, when the overlap determination unit 112 determines that the specific part of the living body and the reflection component do not overlap, the position, angle, size, etc. of the reflection component are configured to be changeable so that the specific part of the living body and the reflection component overlap. When the composite image generation unit 110 includes the perturbation processing unit 111 (FIG. 14) described in the fourth embodiment, the overlap determination unit 112 may output an instruction to the perturbation processing unit 111 so as to perform perturbation processing such that the specific part of the living body and the reflection component overlap. Specifically, an instruction may be issued such that the reflection component overlaps with the specific part of the living body by executing at least one of movement processing, rotation processing, enlargement processing, and reduction processing on the reflection component.

[0122] (Example of overlap determination) Next, with reference to FIG. 18, a specific example will be given and described for the overlap determination of the reflection component by the above-described overlap determination unit 112. FIG. 18 is a conceptual diagram showing an example of determining the overlap between the iris and the reflection component.

[0123] In the example shown in FIG. 18(a), a part of the rectangular region that is the reflection component overlaps with the iris that is the specific part of the living body. In this case, since the reflection component will hide a part of the iris, it becomes difficult to estimate information regarding the iris from the composite image. Therefore, such a composite image is appropriate training data in learning assuming the presence of the reflection component. Therefore, in this case, there is no need to change the position of the reflection component.

[0124] On the other hand, in the example shown in FIG. 18(b), none of the rectangular regions that are the reflection components overlap with the iris that is the specific part of the living body. In this case, since the iris is not hidden by the reflection component, it is relatively easy to estimate information regarding the iris from the composite image. Therefore, such a composite image is inappropriate (not wasted but relatively less effective) training data in learning assuming the presence of the reflection component. Therefore, in this case, it is preferable to change the position of the reflection component so that one of the rectangular regions overlaps with the iris.

[0125] Regarding how much the specific part of the living body and the reflection component should overlap, it may be determined in advance as a hyperparameter. For the specific part of the living body and the reflection component, for example, in some cases, it may be appropriate to make them overlap as much as possible, while in other cases, it may be appropriate to prevent the overlapping part from becoming excessive. The hyperparameter may be set in consideration of such conditions. Also, the degree of overlap between the specific part of the living body and the reflection component may always be constant, or may change each time the synthesis process is performed.

[0126] Note that even when the specific part of the living body and the reflection component do not overlap, simply positioning the reflection component around the specific part of the living body (that is, bringing the specific part of the living body and the reflection component closer to each other) can still obtain the corresponding technical effects described later. Therefore, the overlap determination unit 112 may be configured to determine whether the distance between the specific part of the living body and the reflection component is within a predetermined range, rather than determining whether the specific part of the living body and the reflection component overlap. The predetermined range in this case may be determined in advance with an appropriate value (that is, a value that enables more effective learning). Also, it may be determined probabilistically whether the specific part of the living body and the reflection component overlap. For example, the synthesis of the reflection component may be performed such that there is a 0.5 probability of overlapping and a 0.5 probability of not overlapping. Not only whether it overlaps or not, but also the degree of overlap and the probability of being located around may be determined probabilistically in advance.

[0127] (Operation flow) Next, with reference to FIG. 19, the operation flow of the learning system 10 according to the fifth embodiment will be described. FIG. 19 is a flowchart showing the operation flow of the learning system according to the fifth embodiment. In FIG. 19, the same reference numerals are given to the processes similar to those shown in FIG. 4.

[0128] As shown in FIG. 19, when the learning system 10 according to the fifth embodiment operates, first, the composite image generation unit 110 acquires an image data set of the original image (step S101). Then, training data composed of pairs of the original image and the correct answer is acquired from the acquired image data set by a predetermined number of batch sizes (step S150).

[0129] Subsequently, the overlap determination unit 112 determines whether or not a specific part of the living body and the reflection component overlap, and determines the composite position so that the specific part of the living body and the reflection component overlap (step S501). Then, the composite image generation unit 110 composites the reflection component onto the original image based on the composite position determined by the overlap determination unit 112 (step S102). The composite image generation unit 110 outputs the generated composite image to the iris information estimation unit 120.

[0130] Subsequently, the iris information estimation unit 120 estimates information regarding the iris based on the composite image generated by the composite image generation unit 110 (step S103). Then, the learning unit 130 performs learning processing of the iris information estimation unit 120 based on the estimated information regarding the iris estimated by the iris information estimation unit 120 and the correct answer information in the original image (step S104).

[0131] Subsequently, the learning unit 130 determines whether or not all learning has been completed (step S105). If it is determined that the learning has been completed (step S105: YES), the series of processes ends. On the other hand, if it is determined that the learning has not been completed (step S105: NO), the process may be started again from step S150. As already described, in the second and subsequent steps S501, the composite position may be determined so that the degree of overlap changes from the first time.

[0132] (Technical Effect) Next, the technical effect obtained by the learning system 10 according to the fifth embodiment will be described.

[0133] As described with reference to FIGS. 17 to 19, in the learning system 10 according to the fifth embodiment, the reflection component is synthesized so as to at least partially overlap with a specific part of the living body. By doing so, an image in which a part of the specific part of the living body is hidden by the reflection component can be created, and thus a composite image suitable for learning the iris information estimation unit 120 can be generated. Therefore, it becomes possible to efficiently perform learning for the iris information estimation unit 120.

[0134] <Sixth Embodiment> The learning system 10 according to the sixth embodiment will be described with reference to FIGS. 20 to 22. Note that the sixth embodiment is different from the first to fifth embodiments described above only in some configurations and operations, and the other parts may be the same as those of the first to fifth embodiments. For this reason, hereinafter, descriptions of parts overlapping with the embodiments already described will be omitted as appropriate.

[0135] (Functional Configuration) First, with reference to FIG. 20, the functional configuration of the learning system 10 according to the sixth embodiment will be described. FIG. 20 is a block diagram showing the functional configuration of the learning system according to the sixth embodiment. In FIG. 20, the same reference numerals are given to the same elements as those shown in FIG. 2.

[0136] As shown in FIG. 20, the learning system 10 according to the sixth embodiment includes, as processing blocks for realizing its functions, a composite image generation unit 110, an iris information estimation unit 120, and a learning unit 130. In particular, in the learning system 10 according to the sixth embodiment, the composite image generation unit 110 includes a region number determination unit 113, a region coordinate determination unit 114, a region size determination unit 115, and a region extraction unit 116.

[0137] Note that the reflection component is not input to the composite image generation unit 110 according to the sixth embodiment. The composite image generation unit 110 according to the sixth embodiment does not synthesize the reflection component input from the outside, but determines the region corresponding to the reflection component by the region number determination unit 113, the region coordinate determination unit 114, and the region size determination unit 115, and the region extraction unit 116 cuts out the region, so as to generate a composite image (cut-out image) reflecting the reflection component.

[0138] The region number determination unit 113 is configured to be able to determine the number of regions to be cut out from the original image (that is, the number of regions corresponding to the reflection component). The region number determination unit 113 may be configured to randomly determine the number of regions according to, for example, random numbers. Alternatively, the region number determination unit 113 may be configured to determine the number of regions according to the user's operation. Regarding the number of regions, for example, an appropriate range (for example, 1 to 5, etc.) is set, and the number of regions may be appropriately determined within that range.

[0139] The region coordinate determination unit 114 is configured to be able to determine the coordinates of the region to be cut out from the original image (that is, the position of the region corresponding to the reflection component). The region coordinate determination unit 114 may be configured to randomly determine the region coordinates according to, for example, random numbers. Alternatively, the region coordinate determination unit 114 may be configured to determine the region coordinates according to the user's operation. When there are a plurality of regions (that is, when the number of regions is determined to be 2 or more), the region coordinate determination unit 114 may determine the coordinates for each of the plurality of regions.

[0140] The region size determination unit 115 is configured to be able to determine the size of the region to be cut out from the original image (that is, the size of the region corresponding to the reflection component). The region size determination unit 115 may be configured to randomly determine the region size according to, for example, random numbers. Alternatively, the region size determination unit 115 may be configured to determine the region size according to the user's operation. When there are a plurality of regions (that is, when the number of regions is determined to be 2 or more), the region size determination unit 115 may determine the size for each of the plurality of regions.

[0141] The region extraction unit 116 is configured to be able to execute a process of extracting the region determined by the region number determination unit 113, the region coordinate determination unit 114, and the region size determination unit 115 from the original image. Specific examples of the extraction process will be described in detail below.

[0142] (Example of extraction process) Next, with reference to FIG. 21, the region extraction process by the above-described composite image generation unit 110 (specifically, the region number determination unit 113, the region coordinate determination unit 114, the region size determination unit 115, and the region extraction unit 116) will be specifically described. FIG. 21 is a conceptual diagram showing an example of a method for generating a composite image by extracting from an original image.

[0143] As shown in FIG. 21, in the extraction process, the region corresponding to the reflection component (that is, the region determined by the region number determination unit 113, the region coordinate determination unit 114, and the region size determination unit 115) is extracted from the original image. Note that the region extracted from the original image is a region imitating the reflection component, and thus is not determined based on a specific reflection image or the like. However, the region extracted from the original image may be determined so as to be a reflection component close to an actual reflection image. In the example shown in the figure, a rectangular region is extracted, but the shape is not limited to a rectangle. For example, other polygonal, circular, or elliptical regions may be extracted. When there are a plurality of extraction regions, the coordinates and sizes of the extraction regions may be determined independently for each region.

[0144] The extracted region is filled with specific pixel values close to the actual reflection component. The pixel values of the extracted region may be set, for example, in the value range of 200 to 255. Note that the pixel values of the extracted region may be filled so that all pixels within one region have the same pixel value, or different pixel values may be selected for each pixel.

[0145] Also, when generating a plurality of composite images by the cutting process, the cutting process does not necessarily have to be performed on all the composite images. For example, an image on which the cutting process has been performed may be used for learning, while an image on which the cutting process has not been performed (i.e., an image in its original state) may also be used for learning. Whether to perform the cutting process may be determined according to a predetermined probability. When the cutting process is not performed, the area cutting unit 116 may be made not to execute the process, or the area number determination unit 113 may determine the number of areas as "0".

[0146] The cutout area may be set, for example, to be around the feature points of a specific part of the living body. That is, the cutting process may be performed so as to obtain a composite image similar to the above-described fifth embodiment (i.e., a configuration in which a specific part of the living body and the reflection component overlap). Also, when there are a plurality of feature points of a specific part, the cutting process may be performed after determining whether to perform the cutting process for each feature point. For example, when the iris is a specific part, "iris center", "pupil center", "on the iris circle", "eyelid", etc. may be used as feature points, and for each of these parts, it may be determined whether to perform the cutting process. More specifically, a random number may be generated for each feature point, and whether to perform the cutting process may be determined according to the value of the random number. For example, a random number may be generated between 0 and 1, and when the generated random number is greater than 0.5, the cutting process may be executed, and when it is 0.5 or less, the cutting process may not be executed.

[0147] (Flow of operations) Next, with reference to FIG. 22, the flow of operations of the learning system 10 according to the sixth embodiment will be described. FIG. 22 is a flowchart showing the flow of operations of the learning system according to the sixth embodiment. In FIG. 22, the same reference numerals are given to the same processes as those shown in FIG. 4.

[0148] As shown in FIG. 22, when the learning system 10 according to the sixth embodiment operates, first, the composite image generation unit 110 acquires an image data set of the original image (step S101). Then, training data composed of a pair of the original image and the correct answer is acquired from the acquired image data set in a predetermined number of batch sizes (step S150).

[0149] Subsequently, the region number determination unit 113 determines the number of regions to be cut out (step S601). Subsequently, the region coordinate determination unit 114 determines the coordinates of the regions to be cut out (step S602). Subsequently, the region size determination unit 115 determines the sizes of the regions to be cut out (step S603). Then, the region cutting unit 116 cuts out the determined regions from the original image (step S604). Through such processing, a composite image reflecting the reflection component is generated for the original image. The composite image generation unit 110 outputs the generated composite image to the iris information estimation unit 120.

[0150] Subsequently, the iris information estimation unit 120 estimates information regarding the iris based on the composite image generated by the composite image generation unit 110 (step S103). Then, the learning unit 130 performs learning processing of the iris information estimation unit 120 based on the estimated information regarding the iris estimated by the iris information estimation unit 120 and the correct answer information in the original image (step S104).

[0151] Subsequently, the learning unit 130 determines whether all learning has been completed (step S105). If it is determined that the learning has been completed (step S105: YES), the series of processes ends. On the other hand, if it is determined that the learning has not been completed (step S105: NO), the process may be started again from step S150. As already described, in the second and subsequent executions of step S601, a different number of regions may be determined from the first time. Similarly, in the second and subsequent executions of step S602, different region coordinates may be determined from the first time. In the second and subsequent executions of step S603, different region sizes may be determined from the first time.

[0152] (Technical Effect) Next, the technical effects obtained by the learning system 10 according to the sixth embodiment will be described.

[0153] As described with reference to FIGS. 20 to 22, in the learning system 10 according to the sixth embodiment, a composite image (that is, an image reflecting the reflection component in the original image) is generated by the region cutting process. In this way, a composite image simulating the reflection component can be easily generated using only the original image. That is, a learning image corresponding to the reflection component can be generated without obtaining the reflection component.

[0154] <Seventh Embodiment> The learning system 10 according to the seventh embodiment will be described with reference to FIGS. 23 to 25. Note that the seventh embodiment may be the same as the first to sixth embodiments described above except that some configurations and operations are different, and other parts may be the same as those of the first to sixth embodiments. Therefore, hereinafter, descriptions of parts overlapping with the embodiments already described will be omitted as appropriate.

[0155] (Functional Configuration) First, with reference to FIG. 23, the functional configuration of the learning system 10 according to the seventh embodiment will be described. FIG. 23 is a block diagram showing the functional configuration of the learning system according to the seventh embodiment. In FIG. 23, the same reference numerals are given to the same elements as those shown in FIG. 2.

[0156] As shown in FIG. 23, the learning system 10 according to the seventh embodiment includes a composite image generation unit 110, an iris information estimation unit 120, and a learning unit 130 as processing blocks for realizing its functions. In particular, in the learning system 10 according to the seventh embodiment, the composite image generation unit 110 includes a reflection image generation model 150.

[0157] The reflection image generation model 150 is a model capable of generating a reflection image including a reflection component from a non-reflection image not including the reflection component. Note that the reflection image in the present embodiment is an image used for the learning of the iris information estimation unit 120 and corresponds to the composite image of each of the above-described embodiments. The reflection image generation model 150 is configured as a model learned by a neural network. The specific type of the neural network here is not particularly limited, and an adversarial generation network is mentioned as an example. Below, an example in which the reflection image generation model 150 is learned by an adversarial generation network will be described. The reflection image generation model 150 takes the original image as an input and outputs a reflection image. Further, the reflection image generation model 150 uses a random number as a condition for generating the reflection image. Therefore, the reflection image generation model 150 can output various reflection images according to the input random number.

[0158] (Learning of Reflection Image Generation Model) Next, with reference to FIG. 24, the learning method of the above-described reflection image generation model 150 will be specifically described. FIG. 24 is a block diagram showing a learning method of a reflection image generation model by CycleGAN.

[0159] As shown in FIG. 24, the reflection image generation model 150 is learned using CycleGAN. The configuration during learning by CycleGAN includes the reflection image generation model 150, a reflection image discrimination model 160, a non-reflection image generation model 170, and a non-reflection image discrimination model 180.

[0160] The reflection image generation model 150 takes a non-reflection image (i.e., an image that does not contain a reflection component) and a random number as inputs and generates a reflection image. The reflection image discrimination model 160 takes the reflection image generated by the reflection image generation model 150 and the reflection image of the training data (a reflection image not generated by the reflection image generation model 150) as inputs and outputs the discrimination result. The discrimination result of the reflection image discrimination model 160 is a binary value indicating whether it is real or fake. Note that due to the presence of a random number in the input of the reflection image generation model 150, the reflection image generation model 150 is capable of generating various reflection images. As a result, the reflection components of the reflection images generated by the reflection image generation model 150 are diversified.

[0161] The non-reflection image generation model 170 takes a reflection image (a reflection image not generated by the reflection image generation model 150) as an input and generates a non-reflection image. The non-reflection image discrimination model 180 takes the non-reflection image generated by the non-reflection image generation model 170 and the non-reflection image of the training data (a non-reflection image not generated by the non-reflection image generation model 170) as inputs and outputs the discrimination result. The discrimination result of the non-reflection image discrimination model 180 is a binary value indicating whether it is real or fake.

[0162] In the above CycleGAN, the adversarial loss is calculated and optimized for the reflection image generation model 150 so as to deceive the reflection image discrimination model 160 (make it determine a fake reflection image as a real one). Also, the adversarial loss is calculated and optimized for the reflection image discrimination model 160 so that it is not deceived by the reflection image generation model 150 (does not determine a fake reflection image as a real one). Similarly, the adversarial loss is calculated and optimized for the non-reflection image generation model 170 so as to deceive the non-reflection image discrimination model 180 (make it determine a fake non-reflection image as a real one). Also, the adversarial loss is calculated and optimized for the non-reflection image discrimination model 180 so that it is not deceived by the non-reflection image generation model 170 (does not determine a fake non-reflection image as a real one). At the same time, in addition to the adversarial loss, the generation model uses the reconstruction loss between the generated image obtained after repeatedly estimating the reflection image, non-reflection image, and reflection image and the input image, and further, the reconstruction loss between the generated image and the input image obtained after repeatedly estimating the non-reflection image, reflection image, and non-reflection image. The above learning process may be terminated when a predetermined number of iterations is reached.

[0163] Note that among the above models, only the reflection image generation model 150 is used in the learning system 10, but by simultaneously learning the models related to non-reflection images (i.e., the non-reflection image generation model 170 and the non-reflection image discrimination model 180), it is not necessary to prepare an image pair (a pair of a reflection image and a non-reflection image captured under the same conditions). Specifically, if the reflection image generation model 150, the reflection image discrimination model 160, the non-reflection image generation model 170, and the non-reflection image discrimination model 180 are arranged in a cyclic configuration, it suffices to prepare, as a set of training data, data collected of non-reflection images and data collected of reflection images respectively.

[0164] (Flow of operations) Next, with reference to FIG. 25, the flow of operations of the learning system 10 according to the seventh embodiment will be described. FIG. 25 is a flowchart showing the flow of operations of the learning system according to the seventh embodiment. In FIG. 25, the same reference numerals are given to the same processes as those shown in FIG. 4.

[0165] As shown in FIG. 25, when the learning system 10 according to the seventh embodiment operates, first, the composite image generation unit 110 acquires an image data set of the original image (step S101). Then, training data composed of pairs of the original image and the correct answer is acquired from the acquired image data set in a predetermined number of batch sizes (step S150).

[0166] Subsequently, the reflection image generation model 150 trained by CycleGAN generates a reflection image by taking the original image and a random number as inputs (step S701). The reflection image generation model 150 outputs the generated reflection image to the iris information estimation unit 120.

[0167] Subsequently, the iris information estimation unit 120 estimates information regarding the iris based on the reflection image generated by the reflection image generation model 150 (step S103). Then, the learning unit 130 performs learning processing of the iris information estimation unit 120 based on the estimated information regarding the iris estimated by the iris information estimation unit 120 and the correct answer information in the original image (step S104).

[0168] Subsequently, the learning unit 130 determines whether all learning has been completed (step S105). If it is determined that the learning has been completed (step S105: YES), the series of processes ends. On the other hand, if it is determined that the learning has not been completed (step S105: NO), the process may be started again from step S150. As already described, in step S701, since a random number is input in addition to the original image, the reflection images generated in each iteration of the loop are all different images.

[0169] (Technical Effect) Next, the technical effect obtained by the learning system 10 according to the seventh embodiment will be described.

[0170] As described with reference to FIGS. 23 to 25, in the learning system 10 according to the seventh embodiment, a learning reflection image is generated by the reflection image generation model 150 that has been trained with CycleGAN. In this way, a reflection image can be easily generated using only the original image. That is, a learning reflection image can be generated without obtaining the reflection component.

[0171] <Eighth Embodiment> The estimation system 20 according to the eighth embodiment will be described with reference to FIGS. 26 and 27. Note that the estimation system 20 according to the eighth embodiment is a system including the iris information estimation unit 120 trained in the learning system 10 according to the first to seventh embodiments described above. Its hardware configuration may be the same as that of the learning system 10 according to the first embodiment (see FIG. 1), and other parts may also be the same as those of the learning system 10 according to the first to seventh embodiments. For this reason, hereinafter, descriptions of parts overlapping with the embodiments already described will be omitted as appropriate.

[0172] (Functional Configuration) First, with reference to FIG. 26, the functional configuration of the estimation system 20 according to the eighth embodiment will be described. FIG. 26 is a block diagram showing the functional configuration of the estimation system according to the eighth embodiment. In FIG. 26, the same reference numerals are given to the same elements as those shown in FIG. 2.

[0173] As shown in FIG. 26, the estimation system 20 according to the eighth embodiment includes an iris information estimation unit 120 as a processing block for realizing its functions. The iris information estimation unit 120 according to the eighth embodiment particularly includes an estimation model 121 and an output unit 122.

[0174] The estimation model 121 is a model for estimating information about a specific part of a living body (here, the iris) from the original image. The estimation model 121 is composed of, for example, a neural network and has been trained by the learning system 10 according to the first to seventh embodiments described above. That is, the estimation model 121 is a model in which parameter optimization has been performed using, as training data, an image in which the reflection component is reflected in the original image.

[0175] The output unit 122 is configured to be able to output information about the specific part of the living body estimated by the estimation model 121 to the outside of the system. The output unit 122 may be configured to output information about the specific part of the living body to a biometric authentication system. For example, the output unit 122 may be configured to be able to output information about the feature amount of the iris to an iris authentication system. Note that such an output destination is merely an example, and the output unit 122 may be configured to be able to output information about the specific part of the living body to various other devices and systems.

[0176] (Flow of operations) Next, with reference to FIG. 27, the flow of operations of the estimation system 20 according to the eighth embodiment will be described. FIG. 27 is a flowchart showing the flow of operations of the estimation system according to the eighth embodiment.

[0177] As shown in FIG. 27, when the estimation system 20 according to the eighth embodiment operates, first, an input image including a living body is input to the iris information estimation unit 120 (step S801). For the iris information estimation unit 120, for example, an image captured by a camera may be directly input as it is, or an image stored in a storage or the like may be input.

[0178] Subsequently, the estimation model 121 estimates information about the iris of the living body from the input image (step S802). Then, the output unit 122 outputs the information about the iris of the living body estimated by the estimation model 121 (step S803).

[0179] Note that the above series of processes may be repeatedly executed while the input image is being input. However, when it can be determined that there is no need to estimate information regarding the iris, even if the input image continues to be input, the series of processes may be stopped. For example, when information regarding the iris has been output to the iris authentication system, the series of processes may be terminated when the authentication result by the iris authentication system is obtained.

[0180] (Technical Effect) Next, the technical effect obtained by the estimation system 20 according to the eighth embodiment will be described.

[0181] As described with reference to FIGS. 26 and 27, in the estimation system 20 according to the eighth embodiment, information regarding the iris is estimated using the iris information estimation unit 120 (more specifically, the estimation model 121) learned by the learning system 10 according to the first to seventh embodiments. The learning of the iris information estimation unit 120 is performed using an image in which the reflection component is reflected in the original image, as already described. Therefore, even when the input image includes a reflection component (for example, when the living body wears glasses and the reflection component by the glasses overlaps the iris), it is possible to accurately estimate information regarding a specific part of the living body.

[0182] A program for operating the configuration of the embodiment to realize the functions of the above-described embodiments is recorded on a recording medium, the program recorded on the recording medium is read as code, and a processing method executed by a computer is also included in the scope of each embodiment. That is, a computer-readable recording medium is also included in the scope of each embodiment. In addition, not only the recording medium on which the above program is recorded, but also the program itself is included in each embodiment.

[0183] As the recording medium, for example, a floppy (registered trademark) disk, a hard disk, an optical disk, a magneto-optical disk, a CD-ROM, a magnetic tape, a non-volatile memory card, or a ROM can be used. Further, the present invention is not limited to those that execute processing using only the program recorded on the recording medium, and those that operate on an OS and execute processing in cooperation with other software and the functions of expansion boards are also included in the scope of each embodiment.

[0184] This disclosure can be appropriately changed within a range not contrary to the gist or idea of the invention that can be read from the claims and the entire specification, and a learning system, an estimation system, an estimation method, and a computer program involving such changes are also included in the technical idea of this disclosure.

[0185] <Supplementary Note> Regarding the embodiments described above, it can also be described as follows in the supplementary note below, but is not limited thereto.

[0186] (Supplementary Note 1) The learning system described in Supplementary Note 1 includes a generation means for generating a composite image by synthesizing information regarding a reflection component of light not derived from the original image into the original image including a living body, and a learning means for performing learning of the estimation model based on information regarding a specific part of the living body estimated from the composite image by the estimation model and correct answer information indicating the correct answer of the information regarding the specific part in the original image.

[0187] (Supplementary Note 2) The learning system described in Supplementary Note 2 further includes an extraction means for extracting information regarding the reflection component from the reflection image having the reflection component, and is the learning system described in Supplementary Note 1.

[0188] (Supplementary Note 3) The learning system described in Supplementary Note 3 is characterized in that the extraction means extracts, as information regarding the reflection component, a region where the pixel value in the reflection image is equal to or greater than a predetermined threshold value, and is the learning system described in Supplementary Note 2.

[0189] (Appendix 4) The estimation means, the generation means synthesizes after performing at least one process of movement, rotation, enlargement, and reduction on the information regarding the reflection component, thereby generating a plurality of synthesized images from a single original image. The learning system according to any one of Appendices 1 to 3 is characterized by this.

[0190] (Appendix 5) The learning system according to Appendix 5 is characterized in that the generation means synthesizes the information regarding the reflection component so as to at least partially overlap with the specific part. The learning system according to any one of Appendices 1 to 4 is characterized by this.

[0191] (Appendix 6) The learning system according to Appendix 6 is characterized in that the generation means generates a figure corresponding to the information regarding the reflection component, and generates the synthesized image by cutting out the figure from the original image. The learning system according to any one of Appendices 1 to 5 is characterized by this.

[0192] (Appendix 7) The learning system according to Appendix 7 is characterized in that the generation means generates the synthesized image from the original image using a model learned by a neural network. The learning system according to any one of Appendices 1 to 5 is characterized by this.

[0193] (Appendix 8) The estimation system according to Appendix 8 includes a generation means that synthesizes information regarding a reflection component of light not derived from the original image in the original image including a living body to generate a synthesized image, and a learning means that learns the estimation model based on information regarding a specific part of the living body estimated from the synthesized image by the estimation model and correct information indicating the correct answer of the information regarding the specific part in the original image. The estimation system is characterized by including the estimation model learned using these, and an output means that estimates and outputs information regarding the specific part in the input image using the estimation model.

[0194] (Appendix 9) The learning method described in Supplementary Note 9 is a learning method characterized by generating a composite image by synthesizing information on the reflection component of light not derived from the original image with the original image including a living body, and performing learning of the estimation model based on information on a specific part of the living body estimated from the composite image by the estimation model and correct answer information indicating the correct answer of information on the specific part in the original image.

[0195] (Supplementary Note 10) The computer program described in Supplementary Note 10 is a computer program characterized by causing a computer to operate so as to perform learning of the estimation model based on information on a specific part of the living body estimated from the composite image by the estimation model and correct answer information indicating the correct answer of information on the specific part in the original image, by synthesizing information on the reflection component of light not derived from the original image with the original image including a living body to generate a composite image.

[0196] (Supplementary Note 11) The recording medium described in Supplementary Note 11 is a recording medium characterized by recording the computer program described in Supplementary Note 10.

Explanation of Signs

[0197] 10 Learning system 20 Estimation system 110 Composite image generation unit 111 Perturbation processing unit 112 Overlap determination unit 113 Region number determination unit 114 Region coordinate determination unit 115 Region size determination unit 116 Region extraction unit 120 Iris information estimation unit 121 Estimation model 122 Output unit 130 Learning unit 131 Loss function calculation unit 132 Gradient calculation unit 133 Parameter update unit 140 Reflection component extraction unit 141 Pixel Value Determination Unit 150 Reflected Image Generation Model 160 Reflected Image Discrimination Model 170 Non-Reflected Image Generation Model 180 Non-Reflected Image Discrimination Model

Claims

1. Generating means for generating a composite image by synthesizing information indicating a reflection component of light not derived from the imaging environment of the original image into the original image including a specific part which is a part used for biometric authentication in a living body; Learning means for performing learning of the estimation model so as to improve the accuracy of estimating information indicating the specific part of the living body from the composite image based on information indicating the specific part of the living body estimated from the composite image by an estimation model which is a model for outputting information indicating the specific part using the composite image as an input, and correct answer information indicating the correct answer of the information indicating the specific part in the original image; Comprising; The generating means generates information indicating a plurality of reflection components by executing at least one of movement, rotation, enlargement, and reduction on the information indicating the reflection component a plurality of times, and generates a plurality of composite images by separately synthesizing each of the information indicating the plurality of reflection components into one original image; A learning system characterized by the above.

2. The learning system according to claim 1, further comprising extraction means for extracting information indicating the reflection component from a reflection image having the reflection component.

3. The learning system according to claim 2, wherein the extraction means extracts, as information indicating the reflection component, a region in the reflection image where a pixel value is equal to or greater than a predetermined threshold value.

4. The learning system according to any one of claims 1 to 3, wherein the generating means synthesizes the information indicating the reflection component so as to at least partially overlap the specific part.

5. The generating means generates a figure corresponding to the shape of the region of the reflection component, generates a cutout image by cutting out the figure from the original image, The learning means performs learning using information indicating the specific part of the living body estimated from the cutout image instead of the composite image. The learning system according to any one of claims 1 to 4, characterized by the above.

6. The learning system according to any one of claims 1 to 4, wherein the generating means generates the composite image from the original image using a model which is learned by a neural network and inputs the original image and a random number and outputs the composite image.

7. A generation means for generating a synthetic image by synthesizing information indicating a reflection component of light not derived from the imaging environment of the original image with the original image including a specific part that is a part used for biometric authentication in a living body; based on information indicating the specific part of the living body estimated from the synthetic image by an estimation model that is a model for outputting information indicating the specific part with the synthetic image as an input, and correct answer information indicating the correct answer of the information indicating the specific part in the original image, a learning means for performing learning of the estimation model so as to improve the accuracy of estimating the information indicating the specific part from the synthetic image, and the estimation model learned using the above; Output means for estimating and outputting information indicating the specific part in the input image using the estimation model comprising The generation means generates information indicating a plurality of reflection components by executing at least one of movement, rotation, enlargement, and reduction on the information indicating the reflection component a plurality of times, and generates a plurality of synthetic images by separately synthesizing each of the information indicating the plurality of reflection components with one original image An estimation system characterized by the above.

8. A synthetic image is generated by synthesizing information indicating a reflection component of light not derived from the imaging environment of the original image with the original image including a specific part that is a part used for biometric authentication in a living body, Based on information indicating the specific part of the living body estimated from the synthetic image by an estimation model that is a model for outputting information indicating the specific part with the synthetic image as an input, and correct answer information indicating the correct answer of the information indicating the specific part in the original image, learning of the estimation model is performed so as to improve the accuracy of estimating the information indicating the specific part from the synthetic image, When generating the synthetic image, at least one of movement, rotation, enlargement, and reduction is executed on the information indicating the reflection component a plurality of times to generate information indicating a plurality of reflection components, and a plurality of synthetic images are generated by separately synthesizing each of the information indicating the plurality of reflection components with one original image A learning method characterized by the above.

9. A synthetic image is generated by synthesizing information indicating a reflection component of light not derived from the imaging environment of the original image with the original image including a specific part that is a part used for biometric authentication in a living body, Based on the information indicating the specific part of the living body estimated from the composite image by an estimation model, which is a model that outputs information indicating the specific part using the composite image as input, and the correct answer information indicating the correct answer of the information indicating the specific part in the original image, the learning of the estimation model is performed to improve the accuracy of estimating the information indicating the specific part from the composite image. When generating the composite image, at least one of movement, rotation, enlargement, and reduction is repeatedly executed on the information indicating the reflection component to generate information indicating a plurality of reflection components, and a plurality of composite images are generated by separately synthesizing each of the information indicating the plurality of reflection components with one original image. A computer program characterized by operating a computer as described above.

Images