EP-4738247-A1 - LEARNING IMAGE GENERATION METHOD, LEARNING IMAGE GENERATION DEVICE, LEARNING IMAGE GENERATION PROGRAM, AND LEARNING MODEL GENERATION METHOD

Abstract

This training image generation method includes: a translation step for generating a translated image by translating a source image which includes a target region so as to reduce the discriminability thereof; and an association step for generating a training image in which the translated image and information about the translation target region corresponding to the target region in the translated image are associated with one another as an associative relationship for training a learning model, and using the information about the target region in the source image as translation target region information.

Inventors

• NISHIMORI, Toshiaki
• IENAGA, Kengo
• IEHARA, MASATO

Assignees

• Mitsubishi Heavy Industries Machinery Systems, Ltd.

Dates

Publication Date

20260506

Application Date

20240624

Claims (8)

1. A training image generation method comprising: a translation step of translating a source image including a target region such that identifiability is lowered to generate a translated image; and an association step of generating a training image in which the translated image and information on a translation target region corresponding to the target region in the translated image are associated with each other as a correspondence relationship with which a learning model is trained, information on the target region in the source image being used as the information on the translation target region.
2. The training image generation method according to Claim 1, wherein the translated image is an image of which the identifiability is lower than a threshold value, and the source image is an image of which the identifiability is higher than the threshold value.
3. The training image generation method according to Claim 1, wherein in the translation step, a disturbance is added to the source image to lower the identifiability.
4. The training image generation method according to Claim 1, further comprising: a specifying step of specifying a position and a range of the target region in the source image, wherein in the association step, the position and the range of the target region specified in the specifying step are associated with the translated image as the information on the target region.
5. The training image generation method according to Claim 1, wherein the source image is an image captured by an imaging device.
6. A training image generation device comprising: a generation unit configured to convert a source image including a target region such that identifiability is lowered to generate a translated image; and an association unit configured to generate a training image in which the translated image and information on a translation target region corresponding to the target region in the translated image are associated with each other as a correspondence relationship with which a learning model is trained, information on the target region in the source image being used as the information on the translation target region.
7. A training image generation program causing a computer to execute: a translation process of translating a source image including a target region such that identifiability is lowered to generate a translated image; and an association process of generating a training image in which the translated image and information on a translation target region corresponding to the target region in the translated image are associated with each other as a correspondence relationship with which a learning model is trained, information on the target region in the source image being used as the information on the translation target region.
8. A learning model generation method comprising: a training image generation step of generating a training image by the training image generation method according to any one of Claims 1 to 5; and a training step of training a learning model with the generated training image.

Description

Technical Field The present disclosure relates to a training image generation method, a training image generation device, a training image generation program, and a learning model generation method. Background Art In a case of generating a learning model configured to extract a target region included in an image, a large number of training images in which the image and the target region are associated with each other are required. As a technology for generating the training images, for example, a technology described in PTL 1 is known. Citation List Patent Literature [PTL 1] Japanese Patent No. 7164008 Summary of Invention Technical Problem In a case where the above-described training image is generated, it is required to perform association between the image and the target region with high accuracy. The present disclosure has been made in view of the above-described circumstances, and an object of the present disclosure is to provide a training image generation method, a training image generation device, a training image generation program, and a learning model generation method capable of generating a training image in which accuracy of association between an image and a target region is high as a training image for generating a learning model configured to extract the target region included in the image. Solution to Problem A training image generation method according to the present disclosure includes: a translation step of translating a source image including a target region such that identifiability is lowered to generate a translated image; and an association step of generating a training image in which the translated image and information on a translation target region corresponding to the target region in the translated image are associated with each other as a correspondence relationship with which a learning model is trained, information on the target region in the source image being used as the information on the translation target region. A training image generation program according to the present disclosure causes a computer to execute: a translation process of translating a source image including a target region such that identifiability is lowered to generate a translated image; and an association process of generating a training image in which the translated image and information on a translation target region corresponding to the target region in the translated image are associated with each other as a correspondence relationship with which a learning model is trained, information on the target region in the source image being used as the information on the translation target region. A training image generation device according to the present disclosure includes: a generation unit configured to convert a source image including a target region such that identifiability is lowered to generate a translated image; and an association unit configured to generate a training image in which the translated image and information on a translation target region corresponding to the target region in the translated image are associated with each other as a correspondence relationship with which a learning model is trained, information on the target region in the source image being used as the information on the translation target region. A learning model generation method according to the present disclosure includes: a training image generation step of generating a training image by the training image generation method; and a training step of training a learning model with the generated training image. Advantageous Effects of Invention According to the present disclosure, it is possible to generate a training image in which accuracy of association between an image and a target region is high as a training image for generating a learning model configured to extract a target region included in the image. Brief Description of Drawings FIG. 1 is a functional block diagram illustrating an example of a training image generation device according to an embodiment.FIG. 2 is a view schematically illustrating an example of a source image.FIG. 3 is a view schematically illustrating an example of a translated image.FIG. 4 is a view schematically illustrating an example of a training image.FIG. 5 is a flowchart illustrating an example of a training image generation method according to the present embodiment.FIG. 6 is a view schematically illustrating an example of a case where a learning model is generated based on the training image.FIG. 7 is a flowchart illustrating an example of a learning model generation method according to the present embodiment.FIG. 8 is a view schematically illustrating an example of an operation of the learning model.FIG. 9 is a graph illustrating an example of a ground truth rate of a target region which is output when a captured image is input to the learning model.FIG. 10 is a view illustrating another example of the source image.FIG. 11 is a view illustrating another example of the sourc