Learning image generation method, learning image generation device, learning image generation program and learning model generation method

Abstract

To generate a learning image with high accuracy of association between an image and a target portion as a learning image for generating a learning model that extracts the target portion contained in the image.SOLUTION: A learning image generation method comprises: a transformation step of transforming a reference image that includes a target portion so that its discriminability is reduced to generate a transformed image; and an association step of generating a learning image by associating the transformed image with information of a transformation target portion corresponding to the target portion in the transformed image as a correspondence relationship to be learned by a learning model, and using the information of the target portion in the reference image as the information of the transformation target portion.SELECTED DRAWING: Figure 5

Description

[Technical field]

[0001] The present disclosure relates to a learning image generating method, a learning image generating device, a learning image generating program, and a learning model generating method. [Background technology]

[0002] When generating a learning model for extracting a target portion included in an image, a large number of learning images in which the image and the target portion are associated with each other are required. As a technique for generating learning images, for example, the technique described in Patent Document 1 is known. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent No. 7164008 Summary of the Invention [Problem to be solved by the invention]

[0004] When generating learning images as described above, it is necessary to associate the image with a target portion with high accuracy.

[0005] The present disclosure has been made in consideration of the above, and aims to provide a learning image generation method, a learning image generation device, a learning image generation program, and a learning model generation method that are capable of generating learning images with high accuracy in matching images and target parts as learning images for generating a learning model that extracts target parts contained in an image. [Means for solving the problem]

[0006] The method for generating images for learning according to the present disclosure includes a conversion step of generating a converted image by converting a reference image including a target portion so as to reduce discriminability, and a correspondence step of generating a learning image in which the converted image and information on a conversion target portion in the converted image that corresponds to the target portion as a correspondence relationship for learning by a learning model, and using information on the target portion in the reference image as information on the conversion target portion.

[0007] The learning image generation program of the present disclosure causes a computer to execute a conversion process that converts a reference image including a target portion so as to reduce its distinctiveness to generate a converted image, and a correspondence process that generates a learning image in which the converted image and information on the converted target portion in the converted image that corresponds to the target portion as a correspondence relationship for learning by a learning model, and uses information on the target portion in the reference image as information on the converted target portion.

[0008] The learning image generation device of the present disclosure includes a generation unit that generates a transformed image by transforming a reference image including a target portion so as to reduce distinctiveness, and a correspondence unit that generates a learning image in which the transformed image and information on the transformation target portion in the transformed image that corresponds to the target portion as a correspondence relationship for learning by a learning model, and uses information on the target portion in the reference image as information on the transformation target portion.

[0009] The learning model generation method according to the present disclosure includes a learning image generation step of generating learning images using the above-mentioned learning image generation method, and a learning step of training a learning model based on the generated learning images. Effect of the Invention

[0010] According to the present disclosure, it is possible to generate learning images with high accuracy of association between images and target portions as learning images for generating a learning model that extracts target portions contained in images. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 is a functional block diagram showing an example of a learning image generating device according to the present embodiment. [Diagram 2] FIG. 2 is a diagram illustrating an example of a reference image. [Diagram 3] FIG. 3 is a diagram illustrating an example of a converted image. [Figure 4] FIG. 4 is a diagram illustrating an example of a learning image. [Diagram 5] FIG. 5 is a flowchart showing an example of a learning image generating method according to this embodiment. [Figure 6] FIG. 6 is a diagram illustrating an example of generating a learning model based on learning images. [Figure 7] FIG. 7 is a flowchart showing an example of a learning model generating method according to this embodiment. [Figure 8] FIG. 8 is a diagram illustrating an example of the operation of the learning model. [Figure 9] FIG. 9 is a graph showing an example of the accuracy rate of the target portion output when a captured image is input to the learning model. [Figure 10] FIG. 10 is a diagram showing another example of the reference image. [Figure 11] FIG. 11 is a diagram showing another example of the reference image. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Hereinafter, embodiments of a learning image generating method, a learning image generating device, a learning image generating program, and a learning model generating method according to the present disclosure will be described with reference to the drawings. Note that the present invention is not limited to these embodiments. In addition, the components in the following embodiments include those that are replaceable and easy for a person skilled in the art, or those that are substantially the same.

[0013] 1 is a functional block diagram showing an example of a training image generating device according to the present embodiment. As shown in FIG 1, the training image generating device 100 includes an input / output unit 10, a processing unit 20, and a storage unit 30.

[0014] The input / output unit 10 is an input / output interface that inputs a signal from an external device and outputs a signal to the external device. The input / output unit 10 inputs a reference image (described later) from an external device such as an imaging device. The input reference image can be stored in the storage unit 30, for example. The input / output unit 10 can also output the reference image, a converted image (described later), and a learning image.

[0015] The processing unit 20 performs various information processing. The processing unit 20 includes a processor such as a CPU (Central Processing Unit) and memories such as a ROM (Read Only Memory) and a RAM (Random Access Memory). The processing unit 20 includes a conversion unit 21 and a correspondence unit 22.

[0016] The conversion unit 21 converts a reference image including a target portion so that the discriminability is reduced to generate a converted image. The conversion unit 21 can store the generated converted image in the storage unit 30. The target portion is a part of the reference image. The target portion includes, for example, a portion of the reference image including a predetermined object.

[0017] The reference image may be, for example, an image captured by a photographing device. In this embodiment, the image is not limited to an image captured by a photographing device, and may be data acquired by other types of sensors, such as optical ranging data measured by an optical ranging sensor such as a LiDAR sensor, and radio ranging data measured by a radio ranging sensor such as a RADAR sensor. In this case, each data such as optical ranging data and radio ranging data may be used as data in which measurement values ​​are arranged two-dimensionally, that is, as data equivalent to an image, similar to an image in which pixels are arranged two-dimensionally. In the following description, each data such as optical ranging data and radio ranging data arranged two-dimensionally in this way will be described as being included in the "image".

[0018] The reference image may be, for example, an image with a discriminability higher than a threshold. The reference image may be an image detected under conditions with the least disturbance, such as daytime on a clear day. Disturbances include, for example, weather conditions such as rainfall, snowfall, and fog, and weather conditions such as dawn, dusk, nighttime, and sunlight conditions such as backlighting.

[0019] In this embodiment, the identifiability may be, for example, the visibility of a reference image captured by a photographing device. The visibility may be expressed, for example, by contrast of at least one of the luminance and brightness of the reference image. In addition, in the case of each data such as optical ranging data and radio ranging data, the identifiability may be expressed, for example, by the S / N ratio of the measured value.

[0020] In this embodiment, the reference image may be an image whose discriminability (contrast of at least one of brightness and luminance) is higher than a threshold or equal to or higher than a threshold. The converted image may be an image whose discriminability is lower than the threshold. In this case, the conversion unit 21 converts the reference image whose discriminability is higher than the threshold so that the discriminability is lower than the threshold to generate the converted image. The converted image may be an image whose discriminability is lower than the above threshold.

[0021] The conversion unit 21 can reduce the discriminability by adding, for example, disturbance to the reference image. The conversion unit 21 can convert the reference image using disturbance images such as images showing various weather conditions such as rainfall, snowfall, and fog, and images showing various sunlight conditions such as dawn, dusk, nighttime, and backlight.

[0022] The conversion unit 21 can generate a converted image by learning using a neural network, for example. In this case, when a reference image and a disturbance image are input, the conversion unit 21 can generate a converted image using a technique called Generative Adversarial Networks (GAN) or a technique called Cycle Generative Adversarial Network (Cycle GAN). The conversion unit 21 can also generate a converted image using a technique called a Diffusion Model.

[0023] FIG. 2 is a diagram showing an example of a reference image. The reference image IM1 shown in FIG. 2 is an image of the front taken from a vehicle in a traveling state in the daytime under clear skies. The reference image IM1 includes an image of a preceding vehicle A1 traveling in front of the host vehicle. The reference image IM1 is in a state in which a portion including the image of the preceding vehicle A1 is specified as the target portion P1. When specifying the target portion P1, for example, an operator can look at the reference image IM1 to identify the preceding vehicle A1, and manually set the identified range as the target portion P1. The target portion P1 may be specified by performing image processing on the reference image IM1. The target portion P1 is not limited to the preceding vehicle A1, and may be, for example, another object or person such as a pedestrian, a traffic light, a sign, a motorcycle, or a bicycle.

[0024] FIG. 3 is a diagram showing a schematic example of a converted image. The converted image IM2 shown in FIG. 3 is an image in which a disturbance is added to the reference image IM1 to reduce the discernibility. The converted image IM2 includes an image of a preceding vehicle A2 traveling ahead of the host vehicle. The converted image IM2 may have the same number of pixels as the reference image IM1. In this case, the position and range of the preceding vehicle A2 in the converted image IM2 are the same as the position and range of the preceding vehicle A1 in the reference image IM1.

[0025] The correspondence unit 22 generates a learning image based on the converted image. The correspondence unit 22 can store the generated learning image in the storage unit 30. The learning image is data in which the converted image and information on the target part in the reference image are associated as a correspondence for learning the learning model. In other words, the learning image is data in which the conversion image and information on the target part in the reference image are associated as information on the conversion target part, which is a part of the conversion image that corresponds to the target part of the reference image. When the two-dimensional pixel arrangements of the reference image and the conversion image are the same, the two-dimensional coordinates of each pixel constituting the target part in the reference image and the two-dimensional coordinates of each pixel constituting the conversion target part in the conversion image are the same. In this case, the correspondence unit 22 can associate the position and range of the identified target part with the conversion image as information on the target part.

[0026] FIG. 4 is a diagram showing an example of a learning image. In the learning image IM3 shown in FIG. 4, the converted image IM2 and the information of the target portion P1 in the reference image are associated as a correspondence relationship to be learned by the learning model. The position and range of the conversion target portion P2 in the converted image IM2 are the same as the position and range of the target portion P1 in the reference image IM1. In this embodiment, since the two-dimensional pixel arrangements of the reference image IM1 and the converted image IM2 are the same, the correspondence unit 22 can set the information of the target portion P1 so that the two-dimensional coordinates of each pixel constituting the target portion P1 in the reference image IM1 are the same as the two-dimensional coordinates of each pixel constituting the conversion target portion P2 in the converted image IM2.

[0027] The learning image IM3 generated by the correspondence unit 22 is data in which the information of the target portion P1 is associated with the converted image IM2 with low visibility. For example, if an operator identifies the preceding vehicle A2 by looking at the converted image IM2 and manually sets the identified range as the conversion target portion P2, the preceding vehicle A2 may not be sufficiently identified from the converted image IM2 with low visibility and may be overlooked. For this reason, there is a possibility that a learning image IM3 with insufficient accuracy in the association between the converted image IM2 and the conversion target portion P2 is generated. In contrast, in this embodiment, information on the target portion P1 is identified based on a reference image IM1 with higher visibility than the converted image IM2, and a learning image IM3 in which the information of the target portion P1 is associated with the converted image IM2 as information on the conversion target portion P2 can be generated, so that a learning image IM3 with high accuracy in the association between the converted image IM2 and the conversion target portion P2 can be generated.

[0028] The storage unit 30 stores information such as various programs and data. The storage unit 30 includes storage such as a hard disk drive (HDD) and a solid state drive (SSD). The storage unit 30 can store a reference image input from the input / output unit 10. The storage unit 30 can store a converted image generated by the conversion unit 21 and a learning image generated by the correspondence unit 22.

[0029] The memory unit 30 stores a learning image generation program that causes a computer to execute, for example, a conversion process for converting a reference image including a target portion so as to reduce its distinctiveness to generate a converted image, and a correspondence process for generating a learning image in which the converted image and information on the conversion target portion in the converted image that corresponds to the target portion as a correspondence relationship for learning by a learning model, and which uses information on the target portion in the reference image as information on the conversion target portion.

[0030] In the training image generating device 100, the processor in the processing unit 20 reads out various programs and loads them in memory, thereby executing information processing corresponding to the functions of the above-mentioned units. Examples of the various programs include programs stored in the storage unit 30 and programs recorded on an external recording medium. The training image generating device 100 functions as an information processing device (computer) that executes various information processes. Note that the various programs may be executed by an information processing device other than the training image generating device 100, or the various programs may be executed by the training image generating device 100 and the other information processing device in cooperation with each other.

[0031] 5 is a flowchart showing an example of a method for generating images for training according to the present embodiment. As shown in FIG 5, the method for generating images for training according to the present embodiment includes a specifying step S10, a converting step S20, and a corresponding step S30.

[0032] In the identification step S10, the position and range of the target portion P1 in the reference image IM1 are identified as information on the target portion P1. In the identification step S10, for example, an operator can view the reference image IM1 to identify the preceding vehicle A1, and manually set the identified range as the target portion P1. Note that the target portion P1 may be identified by performing image processing on the reference image IM1.

[0033] In the conversion step S20, a reference image IM1 including the target portion P1 is converted so as to reduce the distinctiveness, thereby generating a converted image IM2.

[0034] In the correspondence step S30, a learning image IM3 is generated in which the converted image IM2 and information on the target portion P1 in the reference image IM1 are associated with each other as a correspondence relationship to be learned by the learning model.

[0035] A learning model can be generated by learning the learning image IM3 generated as described above. Fig. 6 is a diagram showing an example of generating a learning model based on the learning image IM3. The learning processing device 200 shown in Fig. 6 is a computer including a processor such as a CPU (not shown), a memory such as a ROM or RAM (not shown), a storage such as an HDD or SSD (not shown), and an input / output interface (not shown). The learning processing device 200 stores a program, data, etc. for generating a learning model based on the learning image IM3.

[0036] 6, when the learning image IM3 generated as described above is input, learns the input learning image IM3 by machine learning to generate a learning model M. The learning processing device 200 can store the generated learning model M in a storage.

[0037] Fig. 7 is a flowchart showing an example of a learning model generating method according to this embodiment. As shown in Fig. 7, the learning model generating method according to this embodiment includes a learning image generating step S110 and a learning step S120.

[0038] The learning image generating step S110 is a step of generating a learning image IM3, and includes at least the above-described converting step S20 and corresponding step S30 (see FIG. 5), and may also include the identifying step S10 (see FIG. 5).

[0039] In the learning step S120, the generated learning image IM3 is input to the learning processing device 200, and the learning processing device 200 learns the learning image IM3. This makes it possible to generate the learning model M. Note that in the learning step S120, a learning image in which the reference image IM1 and the target portion P1 are associated with each other may be input to the learning processing device 200, and the learning processing device 200 may learn the learning image.

[0040] FIG. 8 is a diagram showing a schematic example of the operation of the learning model M. In FIG. 8, a case will be described where a captured image IM4 including an object A4 is input to the inference processing device 300. The captured image IM4 is an image captured of the front from a vehicle in a traveling state. The captured image IM4 is an image captured in a state where the weather is neither sunny nor daytime, and where various weather conditions such as rainfall, snowfall, and fog occur, or various sunlight conditions such as dawn, dusk, nighttime, and backlight occur. Therefore, the captured image IM4 is an image with lower discriminability (visibility) than the reference image IM1.

[0041] The captured image IM4 and the learning model M generated as described above are input to the inference processing device 300. Based on the learning result of the correspondence between the converted image IM2 with low visibility and the information of the corresponding portion P1 of the reference image IM1, the inference processing device 300 identifies the target portion P4 in the input captured image IM4, i.e., the portion including the object A4 included in the captured image IM4, and outputs the identification result.

[0042] When generating the learning model M, if a learning image in which the reference image IM1 and the target portion P1 are associated is learned, an image captured in daytime under clear skies can be input to the learning processing device 200 as the captured image. In this case, the captured image and the learning model M are input to the inference processing device 300. The inference processing device 300 can identify the target portion in the input captured image based on the learning result of the correspondence between the reference image IM1 with high visibility and the target portion P1 in the reference image IM1, and output the identification result.

[0043] FIG. 9 is a graph showing an example of the accuracy rate of the target part output when a captured image and a learning model M are input to the inference processing device 300. In FIG. 9, the vertical axis indicates the accuracy rate. Also, "All images" on the horizontal axis indicates the results for all captured images input. Also, "High visibility" indicates the results for images of the input captured images with visibility higher than a threshold value. Also, "Low visibility" indicates the results for images of the input captured images with visibility lower than a threshold value.

[0044] 9, the line graph of "Reference image only" shows the accuracy rate when learning model M is generated using only reference image IM1, and the line graph of "Use of converted image" shows the accuracy rate when learning model M is generated using reference image IM1 and learning image IM3.

[0045] By generating a learning model M using the learning image IM3, the learning accuracy of the correspondence between an image (converted image IM2) with visibility lower than the threshold and the target part (conversion target part P2) is improved, and as shown in Figure 9, the accuracy rate increases when a captured image with visibility lower than the threshold is input.

[0046] In addition, since the learning accuracy of images with visibility lower than the threshold is improved, the overall learning accuracy of the learning model M is improved. Therefore, as shown in Fig. 9, the accuracy rate is high not only when a captured image with visibility lower than the threshold is input, but also when a captured image with visibility higher than the threshold is input.

[0047] In the above embodiment, the information on the target portion is, for example, information on a portion of the reference image that includes a predetermined object, but the present invention is not limited to this case. Fig. 10 and Fig. 11 are diagrams showing other examples of the reference image.

[0048] Fig. 10 is a schematic diagram showing an example of an image IM11 recognized by a technique called segmentation. The image IM11 shown in Fig. 10 shows an area partitioned by the outline of a structure or the like. In the image IM11 shown in Fig. 10, information on the target area can be specified by setting each area in the image IM11, such as each part including the partitioned area, or a part including the boundary between the partitioned areas.

[0049] Fig. 11 is a schematic diagram showing an example of an image IM12 recognized by a technique called depth. The image IM12 shown in Fig. 11 shows the depth in three-dimensional directions by shading. In the image IM12, it is possible to specify information on a target portion by setting regions corresponding to the depth, such as a region with the same depth, a region with a rate of change in depth equal to or less than a predetermined value, and a region with a rate of change in depth equal to or more than a predetermined value.

[0050] As described above, the learning image generating method according to the first aspect of the present disclosure includes a conversion step S20 in which a reference image IM1 including a target portion P1 is converted so as to reduce discriminability to generate a converted image IM2, and a conversion target portion matching step S30 in which the converted image IM2 and information on the conversion target portion P2 in the converted image IM2 that corresponds to the target portion P1 in the converted image IM2 are matched as a correspondence for learning by a learning model M to generate a learning image IM3, using information on the target portion P1 in the reference image IM1 as information on the conversion target portion P2.

[0051] According to this configuration, when matching the converted image IM2 with the information of the conversion target portion P2, information of the target portion P1 in the reference image IM1, which has higher visibility than the converted image IM2, is used as the information of the conversion target portion P2, so that a learning image IM3 can be generated that has a high degree of accuracy in matching the converted image IM2 with the conversion target portion P2.

[0052] In the learning image generating method according to the second aspect of the present disclosure, in the first aspect, the converted image IM2 is an image whose discriminability is lower than a threshold, and the reference image IM1 is an image whose discriminability is higher than the threshold.

[0053] According to this configuration, since the converted image IM2 is an image with lower discrimination than the threshold value and the reference image IM1 is an image with higher discrimination than the threshold value, information on the target portion P1 identified in the highly discriminative reference image IM1 can be matched to the converted image IM2 with lower discrimination, thereby generating a learning image IM3 with high accuracy in matching the converted image IM2 with the converted target portion P2.

[0054] In the learning image generating method according to the third aspect of the present disclosure, in the first or second aspect, in the conversion step S20, a disturbance is added to the reference image IM1 to reduce discriminability.

[0055] According to this configuration, it is possible to generate a learning image IM3 that has a high degree of correspondence between a converted image IM2, in which a disturbance is added to a reference image IM1 to reduce discrimination, and the conversion target portion P2.

[0056] The learning image generation method according to the fourth aspect of the present disclosure, in any of the first to third aspects, further includes a determination step S10 for determining the position and range of a target portion P1 in a reference image IM1, and in a correspondence step S30, the position and range of the target portion P1 determined in the determination step S10 is associated with a converted image IM2 as information of the target portion P1.

[0057] According to this configuration, the position and range of the target part P1 are associated with the converted image IM2 as the position and range of the conversion target part P2, so that the target part P1 and the conversion target part P2 can be associated with each other with high accuracy.

[0058] A learning image generating method according to a fifth aspect of the present disclosure is any one of the first to fourth aspects, in which the reference image IM1 is an image captured by an image capture device.

[0059] According to this configuration, it is possible to generate a learning image IM3 that is highly accurately associated with the conversion target portion P2 for a converted image IM2 obtained by converting a reference image IM1 captured by an image capturing device.

[0060] The learning image generating device 100 according to the sixth aspect of the present disclosure includes a conversion unit 21 that converts a reference image IM1 including a target portion P1 so as to reduce discriminability to generate a converted image IM2, and a correspondence unit 22 that generates a learning image IM3 in which the converted image IM2 and information on a conversion target portion P2 in the converted image IM2 that corresponds to the target portion P1 as a correspondence relationship for learning by a learning model M, and uses information on the target portion P1 in the reference image IM1 as information on the conversion target portion P2.

[0061] According to this configuration, when matching the converted image IM2 with the information of the conversion target portion P2, information of the target portion P1 in the reference image IM1, which has higher visibility than the converted image IM2, is used as the information of the conversion target portion P2, so that a learning image IM3 can be generated that has a high degree of accuracy in matching the converted image IM2 with the conversion target portion P2.

[0062] A learning image generation program according to a seventh aspect of the present disclosure causes a computer to execute a conversion process for converting a reference image IM1 including a target portion P1 so as to reduce distinctiveness to generate a converted image IM2, and a correspondence process for generating a learning image IM3 in which the converted image IM2 and information on a conversion target portion P2 in the converted image IM2 that corresponds to the target portion P1 as a correspondence relationship for learning by a learning model M, and which uses information on the target portion P1 in the reference image IM1 as information on the conversion target portion P2.

[0063] According to this configuration, when matching the converted image IM2 with the information of the conversion target portion P2, information of the target portion P1 in the reference image IM1, which has higher visibility than the converted image IM2, is used as the information of the conversion target portion P2, so that a learning image IM3 can be generated that has a high degree of accuracy in matching the converted image IM2 with the conversion target portion P2.

[0064] A learning model generation method according to the eighth aspect of the present disclosure includes a learning image IM3 generation step of generating a learning image IM3 using a learning image generation method according to any one of the first to sixth aspects, and a learning step of training a learning model M to learn the generated learning image IM3.

[0065] According to this configuration, by learning using a learning image IM3 that has high accuracy in matching between the converted image IM2 and the conversion target portion P2, a learning model M can be obtained that can estimate the target area contained in the input image with high accuracy. [Explanation of symbols]

[0066] 10 Input / output section 20 Processing section 21 Conversion section 22 Correspondence Department 30 Storage section 100 Learning image generation device 200 Learning processing device A1, A2 Leading vehicle A4 Object IM1 Reference Image IM2 conversion image IM3 Learning Images IM4 Photo taken IM11,IM12 Images M Learning Model P1,P4 Target part P2 Part to be converted

Claims

1. A transformation step of transforming a reference image including a target portion so as to reduce discriminability to generate a transformed image; a correspondence step of generating a learning image in which the converted image and information on a conversion target portion in the converted image that corresponds to the target portion are associated as a correspondence relationship for learning by a learning model, and the information on the target portion in the reference image is used as the information on the conversion target portion; A method for generating images for learning, comprising:

2. the transformed image is an image in which the discriminability is lower than a threshold value, The reference image is an image in which the discriminability is higher than a threshold value. The method for generating learning images according to claim 1 .

3. In the conversion step, a disturbance is added to the reference image to reduce the discrimination. The method for generating learning images according to claim 1 .

4. The method further includes a step of identifying a position and an area of ​​the target portion in the reference image; In the corresponding step, the position and range of the target part identified in the identifying step are associated with the converted image as information of the target part. The method for generating learning images according to claim 1 .

5. The reference image is an image captured by a photographing device. The method for generating learning images according to claim 1 .

6. A generating unit that generates a converted image by converting a reference image including a target portion so that the discriminability is reduced; a correspondence unit that generates a learning image in which the converted image and information on a conversion target portion in the converted image that corresponds to the target portion are associated as a correspondence relationship for learning by a learning model, and uses information on the target portion in the reference image as the information on the conversion target portion; A learning image generating device comprising:

7. A conversion process for converting a reference image including a target portion so as to reduce discriminability and generating a converted image; a correspondence process for generating a learning image in which the converted image and information on a conversion target portion in the converted image that corresponds to the target portion are associated as a correspondence relationship for learning by a learning model, the correspondence process using information on the target portion in the reference image as information on the conversion target portion; A learning image generation program that causes a computer to execute the above.

8. A learning image generating step of generating learning images by the learning image generating method according to any one of claims 1 to 5; A learning step of learning the generated learning image into a learning model; A learning model generation method comprising:

Images