JP2023065028A - Teacher data production method, image analysis model production method, image analysis method, teacher data production program, image analysis program, and teacher data production device - Google Patents

Abstract

To improve analysis accuracy of an image analysis model.SOLUTION: A teacher data production method which produces teacher data D2 used in machine learning of an image analysis model M for distinguishing a boundary of a region in an image includes: an acquisition step S1 of acquiring learning images D1; a noise provision step S3 of performing processing of providing the learning images D1 with noise to thereby produce noise images; and a teacher data production step S4 of associating the noise images with boundary data showing a boundary of a region in the learning images D1 to produce the teacher data D2.SELECTED DRAWING: Figure 1

Description

Applied for application of Article 30, Paragraph 2 of the Patent Act (No. 1) Date of publication on website March 10, 2020 Website address http://www. sakai-chem. co. jp/jp/ http://www.jp/jp/ sakai-chem. co. jp/jp/kaiseki/index. html

Applied for application of Article 30, Paragraph 2 of the Patent Law (No. 2) Date of publication on website March 12, 2020 Website address http://www. sakai-chem. co. jp/jp/ http://www.jp/jp/ sakai-chem. co. jp/jp/kaiseki/index. html

Applied for application of Article 30, Paragraph 2 of the Patent Law (No. 3) Date of publication on website March 10, 2020 Website address http://www. sakai-chem. co. jp/jp/ http://www.jp/jp/ sakai-chem. co. jp/jp/kaiseki/index. html http://www. sakai-chem. co. jp/jp/kaiseki/pdf/kaiseki_pdf01. pdf

Applied for application of Article 30, Paragraph 2 of the Patent Law (No. 4) Date of posting on website June 14, 2021 Website address https://www. youtube. com/watch? v=GPYbAVEQrUI

Application for application of Article 30, Paragraph 2 of the Patent Act (Part 5) Exhibition date May 19, 2021 to May 21, 2021 Exhibition name CITE JAPAN 10th Cosmetic Industrial Technology Exhibition Venue Pacifico Yokohama (Yokohama, Kanagawa Prefecture) 1-1-1 Minatomirai, Nishi-ku)

Applied for application of Article 30, Paragraph 2 of the Patent Law (No. 6) Posting date of website June 16, 2021 to June 18, 2021 Website address https://www. japan-it-online. jp/ja-jp. html https://www. japan-it-online. jp/ja-jp/about/ai. html

Applied for application of Article 30, Paragraph 2 of the Patent Law (Part 7) Date of publication on website August 4, 2021 to September 3, 2021 (Republished) October 22, 2021 to November 26, 2021 Website address https://www. matsumoto-trd. co. jp/https://www. matsumoto-trd. co. jp/2021exhibition/entry/https://www. matsumoto-trd. co. jp/2021exhibition/greeting/ https://www. matsumoto-trd. co. jp/2021-exhibition/sak. html

Applied for application of Article 30, Paragraph 2 of the Patent Act (Part 8) Date of publication on website August 4, 2021 to September 3, 2021 (Republished) October 22, 2021 to November 26, 2021 Website address https://www. matsumoto-trd. co. jp/https://www. matsumoto-trd. co. jp/2021exhibition/entry/https://www. matsumoto-trd. co. jp/2021exhibition/greeting/ https://www. matsumoto-trd. co. jp/2021exhibition/free presentation/

There is an application for application of Article 30, Paragraph 2 of the Patent Act (Part 9) Exhibition date October 6, 2021 to October 8, 2021 Exhibition name Food Development Exhibition 2021 Venue Tokyo Big Sight (3-Ariake, Koto-ku, Tokyo) 10-1)

There is an application for the application of Article 30, Paragraph 2 of the Patent Act (No. 10) Website publication date September 29, 2021 to November 12, 2021 Website address https://www. powtex. com/online/https://powtex. com/online/expo/exhibitor/5-b13/

Patent Law Article 30, Paragraph 2 Application Applied (Part 11) Exhibition Date October 13, 2021 to October 15, 2021 Exhibition Name International Powder Industry Exhibition Osaka 2021 Venue Intex Osaka (Suminoe, Osaka City, Osaka Prefecture) 1-5-102 Nankohoku, Ward)

There is an application for the application of Article 30, Paragraph 2 of the Patent Law (No. 12) Website publication date October 18, 2021 to October 29, 2021 Website address https://www. chemmate. jp/online https://www. chemmate. jp/online/exhibitor https://www. chemmate. jp/online/exhibitor/list/5 https://www. chemmate. jp/online/exhibitor/booth/472

There is an application for the application of Article 30, Paragraph 2 of the Patent Act (No. 13) Date of publication on website October 18, 2021 to October 29, 2021 Website address https://www. chemmate. jp/online https://www. chemmate. jp/ebook/guidebook2021/book/#target/page_no=1 https://www. chemmate. jp/ebook/guidebook2021/book/#target/page_no=25 https://www. chemmate. jp/online/exhibitor/booth/472

The present invention relates to techniques for analyzing images, and more particularly to techniques for determining boundaries of regions in images.

Image analysis is often used to judge the quality and grade of goods and products, but in order to recognize the target object, the boundary of the area in the image is determined and the target area is distinguished by color etc. A process called binarization (segmentation) is performed. Conventional binarization was performed mainly on a rule-based basis based on grayscale and edge detection, but it is vulnerable to noise that occurs in pixel units and broken areas, and images that can be automatically binarized are limited. rice field. Therefore, for images that cannot be automatically binarized, binarization is performed manually, which takes a huge amount of labor and time. For this reason, it has been difficult to secure an amount of data that can statistically guarantee reliability, and it has not been possible to digitize image data and use it as reliable data.

On the other hand, a method of binarization by artificial intelligence has been proposed instead of a rule-based algorithm.

For example, Patent Literature 1 describes a method of detecting the number and position of cells using machine learning. Patent Document 2 describes a method of detecting an object by predicting a plurality of fused feature maps from an image to be processed using a neural network. In Patent Document 3, a first network learns images of roads and non-roads, predicts road portions and outputs weak segmentations, and then learns weak segmentations in a second network to predict roads. method is described.

JP 2019-91307 A Japanese Patent Publication No. 2020-509488 JP 2019-61658 A

However, in the methods described in Patent Documents 1 to 3, none of the teacher data for machine learning is customized, so even if a sufficient amount of teacher data is prepared, the learned teacher data If the image quality of the image to be analyzed varies, there is a problem that it is difficult to improve the accuracy of binarization.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and an object of the present invention is to improve the analysis accuracy of an image analysis model.

A teacher data generation method according to the present invention is a teacher data generation method for generating teacher data used in machine learning of an image analysis model for determining boundaries of regions in an image, and includes an acquisition step of acquiring a learning image. a noise adding step of generating a noise image by performing a process of adding noise to the learning image; and a training data generation step for generating

According to a preferred embodiment, the processing is at least one of blurring, resizing, cropping, brightness processing, contrast processing and horizontal flipping processing.

According to a preferred embodiment, said processing is a blurring processing.

According to a preferred embodiment, said processing is a brightness processing.

According to a preferred embodiment, said processing is blurring and brightness processing.

According to a preferred embodiment, said processes are blurring, brightness, resizing and contrasting.

According to a preferred embodiment, in the noise adding step, the strength of the noise added to each learning image is random.

The image analysis model generation method according to the present invention performs machine learning using teacher data generated by the above-described teacher data generation method. A learning step is provided for generating an image analysis model that outputs data indicating boundaries.

An image analysis method according to the present invention includes an acquisition step of acquiring an analysis target image; a step;

A training data generation program according to the present invention causes a computer to execute each step of the training data generation method described above.

An image analysis program according to the present invention causes a computer to execute each step of the above image analysis method.

A teacher data generation device according to the present invention is a teacher data generation device that generates teacher data used for learning an image analysis model for determining boundaries of regions in an image, and comprises: an acquisition unit that acquires a learning image; a noise adding unit that generates a noise image by performing a process of adding noise to the learning image; and a training data generation unit that generates the training data.

According to the present invention, it is possible to improve the analysis accuracy of the image analysis model.

1 is a block diagram showing a schematic configuration of an image analysis model generation system according to one embodiment of the present invention; FIG. (a) to (c) are examples of learning images. 2A to 2C are examples of boundary data generated from the learning images shown in FIGS. 2A to 2C, respectively. 1 is a block diagram of an image analysis device according to one embodiment of the present invention; FIG. It is an example of an image to be analyzed. It is an example of a binarized image. It is an example of a noise image. 4 is a flow chart showing a processing procedure of an image analysis model generation method according to one embodiment of the present invention; (a) to (j) are images to be analyzed used in the embodiment of the present invention. 3 is a binarized image generated in an embodiment of the present invention; 3 is a binarized image generated in an embodiment of the present invention; 3 is a binarized image generated in an embodiment of the present invention; 3 is a binarized image generated in an embodiment of the present invention;

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In addition, this invention is not limited to the following embodiment.

(overall structure)
FIG. 1 is a block diagram showing a schematic configuration of an image analysis model generation system 1 according to one embodiment of the present invention. The image analysis model generation system 1 is a system for generating an image analysis model for determining boundaries of regions in an image by the image analysis model generation method according to the present embodiment. a device 3;

(Teacher data generation)
The teacher data generation device 2 is a device that generates teacher data used for machine learning of an image analysis model for determining boundaries of regions in an image by the teacher data generation method according to the present embodiment. The training data generation device 2 can be configured with a general-purpose computer, and includes a processor such as a CPU or GPU, a main storage device such as DRAM or SRAM (not shown), and auxiliary devices such as HDD or SSD. A storage device 20 is provided. The auxiliary storage device 20 stores a plurality of learning images D1 and various programs for operating the teacher data generating device 2, such as a teacher data generating program.

The learning image D1 is an image that serves as a sample for generating teacher data, and may be any image that includes a plurality of areas within the image. FIGS. 2A to 2C are examples of the learning image D1. Each image contains a large number of particles. Although the number of learning images D1 is not particularly limited, it is preferable that the number is sufficient for machine learning.

The teacher data generation device 2 also includes an acquisition unit 21, a boundary data generation unit 22, a noise addition unit 23, and a labeling unit 24 as functional blocks. In this embodiment, each of these units is implemented in software by the processor of the training data generation device 2 reading out a training data generation program into the main memory and executing it.

The acquisition unit 21 is a functional block that acquires the learning image D1. In this embodiment, the acquisition unit 21 acquires the learning image D1 by reading it from the auxiliary storage device 20, but it may acquire the learning image D1 via a communication line with the outside or a recording medium.

The boundary data generator 22 is a functional block that generates boundary data indicating boundaries of regions in the learning image D1. In the present embodiment, the boundary data generation unit 22 displays each area in the learning image D1 in white and the other areas in black according to the user's operation via an input device (not shown). Create boundary data. The boundary data are correct data indicating boundaries of regions in the learning image D1.

FIGS. 3A to 3C are examples of boundary data generated from the learning images D1 shown in FIGS. 2A to 2C, respectively. In the boundary data shown in FIG. 3(a), the portion defined by the fine particles in the surface layer portion with relatively high brightness is displayed in white, and the other portions (boundary of the region, fine particles other than the surface layer portion) are displayed in black. be done. In the boundary data shown in FIG. 3B, the portions corresponding to the rectangular particles are displayed in white, and the other portions are displayed in black. In the boundary data shown in FIG. 3(c), portions corresponding to particles of a certain size or larger are displayed in white, and other portions are displayed in black. Note that the form of the boundary data is not limited to this. For example, it may be data in which only the boundary of the area is displayed with a line and the other portion is filled with white or the like.

The noise adding unit 23 shown in FIG. 1 is a functional block that generates a noise image by performing a process of adding noise to the learning image D1. Details of the process of adding noise will be described later.

The labeling unit 24 is a functional block that associates the noise image added with noise by the noise adding unit 23 and the boundary data generated by the boundary data generating unit 22 to generate teacher data D2. The teacher data D2 is accumulated in the auxiliary storage device 20 and then transferred to the machine learning device 3 as a learning data set.

(machine learning)
The machine learning device 3 is a device that generates an image analysis model M by machine learning. The machine learning device 3 can be configured with a general-purpose computer, and, like the training data generation device 2, includes a processor such as a CPU or GPU, a main storage device such as DRAM or SRAM (not shown), and an HDD or SSD. and an auxiliary storage device 30 such as The auxiliary storage device 30 stores teacher data D2 transferred from the teacher data generation device 2, various programs for operating the machine learning device 3, and the like.

Although the machine learning device 3 is provided on the cloud in this embodiment, it may be installed in the same place as the training data generation device 2, or alternatively, the training data generation device 2 and the machine learning device 3 may be installed in the same place. may be configured integrally.

The machine learning device 3 includes a learning section 31 as a functional block. By performing machine learning based on the teacher data D2, the learning unit 31 generates an image analysis model M that outputs data indicating boundaries of regions in the analysis target image when the analysis target image is input. Although the learning method is not particularly limited, for example, deep learning, support vector machine, random forest, etc. can be used. The generated image analysis model M is stored in the auxiliary storage device 30 and then used for image analysis.

(image analysis)
FIG. 4 is a block diagram of the image analysis device 4 according to this embodiment. The image analysis device 4 is a device used for executing an image analysis method for analyzing an image to be analyzed, and can be configured by a general-purpose computer. The image analysis device 4 includes, as a hardware configuration, a processor such as a CPU or GPU, a main storage device (not shown) such as DRAM or SRAM, and an auxiliary storage device 40 such as HDD or SSD.

In the auxiliary storage device 40, in addition to the image analysis model M and the analysis target image D3 generated by the image analysis model generation system 1 (image analysis model generation method according to the present embodiment) shown in FIG. Various programs for operating the image analysis device 4 are stored. The image D3 to be analyzed is, for example, an image that has been sent to the image analysis apparatus 4 from a requester for image analysis.

The image analysis device 4 includes an image analysis section 41 and a binarization processing section 42 as functional blocks. In this embodiment, each of these units is implemented in software by the processor of the image analysis device 4 reading an image analysis program into the main storage device and executing it.

The image analysis unit 41 is a functional block that uses the image analysis model M to determine boundaries of regions in the analysis target image D3. In this embodiment, the image analysis unit 41 reads the analysis target image D3 and the image analysis model M from the auxiliary storage device 40 to the main storage device. FIG. 5 is an example of the analysis target image D3.

Subsequently, the image analysis unit 41 inputs the analysis target image D3 to the image analysis model M, and in response, the image analysis model M outputs data indicating boundaries of regions in the analysis target image D3. The data is an image in which the portion corresponding to the area is displayed in white and the other portion is displayed in black, as shown in FIG. Note that the data is not limited to black and white, and may be binary data instead of an image. The image analysis unit 41 inputs the output data of the image analysis model M to the binarization processing unit 42 .

Based on the data output by the image analysis unit 41, the binarization processing unit 42 recognizes regions divided by the determined boundaries in the image D3 to be analyzed, and colors or labels the recognized regions, or labels them. By displaying them simultaneously, a binarized image D4 is generated. The binarized image D4 is stored in the auxiliary storage device 40, and then sent as the analysis result to the client of the image analysis.

The boundary data of the image analysis unit 41 is not limited to binarization, but may be ternary or more boundary data. It is possible.

FIG. 6 is an example of the binarized image D4. In the binarized image D4, the recognized area is colored. In the image analysis device 4, the binarization processing unit 42 is not an essential component, and output data of the image analysis model M, that is, data similar to the boundary data shown in FIG. 3 may be used as the analysis result.

(noise addition processing)
In the conventional method, boundary data are labeled without imparting noise to learning images when generating teacher data. Therefore, for example, when there is variation in the image quality of the image to be analyzed with respect to the learned teacher data, there is a problem that the accuracy of recognizing the area in the image to be analyzed by the image analysis model is low. Therefore, in the present embodiment, in order to improve the recognition rate of the area in the binarized image D4, the noise adding section 23 in the teacher data generation device 2 adds noise to the learning image D1.

Noise addition processing by the noise addition unit 23 of the training data generation device 2 will be described below. There are six types of noise addition processing in this embodiment: blur processing (Blur), resize processing (Rsize), crop processing (Crop), brightness processing (Brightness), contrast processing (Contrast), and left-right flip processing (Hflip). However, the present invention is not limited to these.

The blurring process is a process of reducing the resolution of the image by applying a blurring filter to the learning image D1. Resize processing is processing for reducing the size of an image. Crop processing is processing for cutting out part of an image. Brightness processing is processing that changes (decreases or increases) the brightness of an image. Contrast processing is processing that alters (reduces or increases) the contrast of an image. The horizontal reversal process is a process for horizontally reversing the orientation of an image. The noise adding unit 23 generates a noise image by performing at least one of these processes on one learning image D1.

FIG. 7 is an example of a noise image. This noise image is obtained by subjecting the learning image D1 shown in FIG. 2A to left-right reversal processing, blurring processing, contrast processing (lowering contrast), brightness processing (increasing brightness), cropping processing, and resizing processing in this order. It was generated by

The noise image is associated with the boundary data by the labeling unit 24 to generate the teacher data D2, and the learning in the machine learning device 3 generates the image analysis model M. FIG. Since the image analysis model M generated in this way is learned using a noise image to which noise is added, if there is variation in the image quality of the image to be analyzed with respect to the learned teacher data However, the analysis target image can be analyzed with high accuracy.

All of the six types of noise processing may be performed on one learning image D1, or noise processing relating to a single noise or a combination of a plurality of noises may be performed. For example, as shown in the examples below,
・Blurring processing only ・Brightness processing only ・Blurring processing and brightness processing only ・Even if only blurring processing, brightness processing, resizing processing, and contrast processing are performed, a sufficiently high-precision image analysis model can be generated.

Moreover, it is preferable to randomly change the strength of the noise to be added to each learning image D1. That is, the noise adding unit 23 may randomly change the strength of the noise to be added for each learning image D1. By learning the image analysis model M using the noise image generated in this way, it is possible to generate an image analysis model that can analyze with high accuracy even an image to be analyzed that has variations in image quality. .

(Processing procedure)
The processing procedure of the image analysis model generation method according to this embodiment will be described below using the flowchart shown in FIG. The image analysis model generation method is executed by the image analysis model generation system 1 shown in FIG.

In step S<b>1 (acquisition step), the acquisition unit 21 of the teacher data generation device 2 acquires the learning image D<b>1 from the auxiliary storage device 20 . Subsequently, in step S2, the boundary data generation unit 22 generates boundary data indicating boundaries of regions in the learning image D1. A noise image is generated by performing a process of adding . The order of steps S2 and S3 is not particularly limited.

Subsequently, in step S4 (teaching data generating step), the labeling unit 24 associates the noise image with the boundary data to generate teaching data D2.

Through steps S1 to S4, one piece of teacher data D2 is generated from one learning image D1. Steps S1 to S4 are repeated until all learning images D1 are processed (YES in step S5), and then the process proceeds to step S6 (learning step).

In step S6, the learning unit 31 of the machine learning device 3 generates the image analysis model M by performing machine learning based on the teacher data D2.

(Additional notes)
Although the embodiments of the present invention have been described above, the present invention is not limited to the above embodiments, and various modifications are possible without departing from the scope of the invention. For example, in the above-described embodiment, the noise addition processing by the noise addition unit 23 includes six types of noise addition processing such as blurring processing. processing), drop processing (processing to remove any size and number part of the image), rotation processing (processing to rotate the image at any angle), color processing (processing to change the color of the image within any range) and so on.

Specific examples are given below to describe the present invention in detail, but the present invention is not limited thereto.

(Generation of image analysis model)
In this embodiment, 97 learning images common to each model are subjected to different noise processing (at least one of blur processing, resizing processing, crop processing, brightness processing, contrast processing, and left-right inversion processing) to obtain a noise image. is generated and associated with boundary data prepared in advance to generate teacher data, and machine learning (deep learning) is performed using the teacher data to generate 15 image analysis models M1 to M15. In addition, as a comparative example, teacher data is generated by associating boundary data with the above learning image that has not undergone noise processing, machine learning (deep learning) is performed using the teacher data, and an image analysis model M0 is generated. generated.

Tables 1 and 2 show the presence or absence or strength of noise given to the learning image D1 used for learning the image analysis models M1 to M15 and M0.

In each of the above tables, for each noise of resizing, cropping, and left/right inversion, only presence/absence of addition (Yes/No) is shown, and the strength of each noise of resizing and cropping is constant. In the resizing process, both the vertical and horizontal dimensions of the image were reduced by 0.5 times. In the cropping process, it was cut to a size of 0.5×0.5.

Blurring, brightness, and contrast noise were applied with different strengths for each model.

The intensity of blurring was controlled by the number of times the training image was filtered. "Low" means once, and "Middle" means twice. An averaging filter was used for the blurring filter, and the kernel (range of values for averaging) was 5×5. "Arbitrarily" means that the number of times the blurring filter is applied is arbitrarily changed within the range of 1 to 2 depending on the learning image.

Regarding brightness, if the original learning image is 1, "Low" is in the range of 0.8 to 1.2, "Middle" is in the range of 0.5 to 1.5, and "High" is in the range of 0.5 to 1.5. It ranges from 0.3 to 1.7, and "Higher" means randomly varying brightness for each image, ranging from 0.1 to 2.0.

Regarding the contrast, if the original learning image is 1, "Middle" is in the range of 0.5 to 1.5, "High" is in the range of 0.3 to 1.7, and "Higher" is It means that the contrast is randomly changed for each image within the range of 0.1 to 2.0.

(image analysis)
Subsequently, 10 images A to J to be analyzed shown in FIGS. 9A to 9J are input to the image analysis models M1 to M15 and M0, and the output data from each image analysis model is binarized. generated a binarized image.

10 to 13 show binarized images obtained by binarizing the data obtained by analyzing the images A to J to be analyzed by the image analysis models M1 to M15 and M0. In each binarized image, portions recognized as regions are colored.

Tables 3 and 4 show the number of recognized regions in the binarized images corresponding to the image analysis models M1 to M15 and M0.

Furthermore, Tables 5 and 6 show the ratio of the number of recognitions of the image analysis models M1 to M15 to the number of recognitions of the image analysis model M0.

All of the image analysis models M1 to M15 according to the present example tended to improve the number of recognitions compared to the image analysis model M0 according to the comparative example. Therefore, it can be seen that the analysis accuracy of the image analysis model can be improved by performing noise processing on the learning image. In particular, from the results of image analysis models M11 to M14, noise processing is
・Blurring processing only ・Brightness processing only ・Blurring processing and brightness processing only .

Further, when the image analysis models M1 to M3 are compared, Low or Arbitrarily is effective for blur intensity. When comparing the image analysis models M4 to M6, Low or High is effective for brightness. ˜M9, it can be seen that No or Higher contrast is effective.

In addition, the image analysis models M1 to M4, M6, M7, M9, and M12 to M15 are found to have high generalization performance (strength/weakness in analysis depending on image type).

1 Image analysis model generation system 2 Teacher data generation device 20 Auxiliary storage device 21 Acquisition unit 22 Boundary data generation unit 23 Noise addition unit 24 Labeling unit 3 Machine learning device 30 Auxiliary storage device 31 Learning unit 4 Image analysis device 40 Auxiliary storage device 41 Image analysis unit 42 Binary processing unit D1 Learning image D2 Teacher data D3 Analysis target image D4 Binary image M Image analysis model

Claims (12)

A teacher data generation method for generating teacher data used in machine learning of an image analysis model for determining boundaries of regions in an image,
an acquisition step of acquiring training images;
a noise adding step of generating a noise image by performing a process of adding noise to the learning image;
a training data generation step of generating training data by associating the noise image with boundary data indicating boundaries of regions in the learning image;
A teacher data generation method comprising: 2. The teaching data generation method according to claim 1, wherein said processing is at least one of blurring processing, resizing processing, cropping processing, brightness processing, contrast processing, and horizontal reversing processing. 3. The training data generation method according to claim 2, wherein said processing is blur processing. 3. The teaching data generating method according to claim 2, wherein said processing is brightness processing. 3. The teaching data generation method according to claim 2, wherein said processing is blur processing and brightness processing. 3. The teaching data generating method according to claim 2, wherein said processing is blur processing, brightness processing, resizing processing, and contrast processing. 7. The teacher data generation method according to claim 1, wherein in said noise adding step, the strength of said noise added to each learning image is random. By performing machine learning using the teacher data generated by the teacher data generation method according to any one of claims 1 to 7, when an image to be analyzed is input, the boundary of the area in the image to be analyzed An image analysis model generation method comprising a learning step of generating an image analysis model that outputs data representing: an acquisition step of acquiring an image to be analyzed;
A determination step of determining boundaries of regions in the image to be analyzed using the image analysis model generated by the image analysis model generation method according to claim 8;
An image analysis method comprising: A teacher data generation program that causes a computer to execute each step of the teacher data generation method according to any one of claims 1 to 7. An image analysis program that causes a computer to execute each step of the image analysis method according to claim 10. A training data generation device for generating training data used for learning an image analysis model for determining boundaries of regions in an image,
an acquisition unit that acquires a learning image;
a noise adding unit that generates a noise image by performing a process of adding noise to the learning image;
a training data generation unit that generates training data by associating the noise image with boundary data indicating boundaries of regions in the learning image;
A teacher data generation device comprising:

Images