Learned model generation method, learned model, surface defect inspection method, steel manufacturing method, pass/fail determination method, grade determination method, surface defect determination program, pass/fail determination program, determination system, and steel manufacturing equipment

Patent metadata

Field	Value
Publication number	US-12266094-B2
Application number	US-201917413759-A
Country	US
Kind code	B2
Filing date	Oct 31, 2019
Priority date	Dec 25, 2018
Publication date	Apr 1, 2025
Grant date	Apr 1, 2025

Abstract

A learned model generation method includes: using a teacher image including a defect map that is an image indicating a distribution of a defect portion of a surface of steel and having an equal image size, and presence/absence of periodic defects assigned in advance to the defect map; and generating a learned model by machine learning, the learned model for which: an input value is a defect map that is an image indicating a distribution of a defect portion of a surface of steel and having an image size of the equal image size; and an output value is a value concerning presence/absence of periodic defects in the defect map.

First claim

The invention claimed is:

1. A method comprising:
acquiring imaged data for which a surface of steel has been captured;
creating a defect map that is an image indicating a distribution of a defect portion based on the imaged data;
creating a determination image for each defect map by converting an image size of the defect map into a predetermined image size; and
inputting the determination image into a learned model, outputting a value concerning presence/absence of periodic defects, and determining the presence/absence of periodic defects for the determination image based on the value concerning the presence/absence of periodic defects, wherein
the learned model is trained using
a teacher image as an input, the teacher image (i) having the predetermined image size and (ii) being a defect map indicating a distribution of defect portions on the surface of the steel, and
a value concerning presence or absence of periodic defects in the teacher image as an output, and

the predetermined image size has a correct answer rate associated with the learned model that falls within an allowable range.

2. A steel manufacturing method comprising:
detecting periodic defects on a surface of a steel using the method according to claim 1 ; and
manufacturing steel by controlling a manufacturing condition based on the detection result.

3. The method according to claim 1 , further comprising:
calculating a defect mixing rate of the determination image based on the defect portion; and
determining a grade of the steel corresponding to the determination image, based on the calculated defect mixing rate and a determination result of the presence/absence of periodic defects for the determination image.

4. The method according to claim 3 , further including a steel manufacturing sorting step of sorting steel for each grade.

5. The method according to claim 1 ,
when an image size of the defect map used as the teacher image is different from the predetermined image size, the image size of the defect map is converted into the predetermined image size to use the defect map as the input value.

6. The method according to claim 5 , wherein the learned model is being generated by using a test image including a defect map indicating a distribution of a defect portion of a surface of steel and having the equal image size as the teacher image, and presence/absence of periodic defects in the defect map, wherein generating the learned model by using the test image includes:
inputting the defect map of the test image into the learned model to output a value concerning presence/absence of periodic defects in the test image;
determining the presence/absence of periodic defects in the test image based on the value concerning the presence/absence of periodic defects;
calculating a correct answer rate for the test image by comparing the determined presence/absence of periodic defects with the presence/absence of periodic defects in the defect map of the test image; and
adjusting a generation condition of the learned model according to the calculated correct answer rate.

7. The method according to claim 6 , wherein adjusting the generation condition of the learned model is to change the equal image size to another equal image size of different image size.

8. The method according to claim 6 , wherein generating the learned model further includes:
setting a plurality of types of image sizes different from one another as the predetermined image size;
calculating the correct answer rate for each type of the image size by using a pair of the teacher image and the test image of each of the plurality of types of image sizes; and
generating the learned model by using the teacher image and the test image that have the image size having a highest calculated correct answer rate, the highest calculated correct answer rate being the correct answer rate of the predetermined image size.

9. The method according to claim 1 , wherein the learned model includes:
an input layer configured to be input, as a determination image, a defect map indicating a distribution of a defect portion of a surface of steel and having the predetermined image size,
an output layer configured to output a value concerning presence/absence of periodic defects for the determination image,
an intermediate layer for which parameters have been learned by using a teacher image, the teacher image for which:
an input is a defect map having the predetermined image size as an image size of the determination image; and
an output is a value concerning the presence/absence of periodic defects for the defect map,

the learned model causes a computer to function to:
perform calculation on the determination image of the predetermined image size input into the input layer based on the parameters, and
output the value concerning the presence/absence of periodic defects from the output layer.

10. A method comprising:
acquiring imaged data for which a surface of steel has been captured;
creating a defect map that is an image indicating a distribution of a defect portion based on the imaged data;
creating a determination image for each defect map by converting an image size of the defect map into a predetermined image size; and
inputting the determination image into a learned model, outputting a value concerning presence/absence of periodic defects, and determining a pass/fail of the steel based on the value concerning the presence/absence of periodic defects, wherein
the learned model is trained using
a teacher image as an input, the teacher image (i) having the predetermined image size and (ii) being a defect map indicating a distribution of defect portions on the surface of the steel, and
a value concerning presence or absence of periodic defects in the teacher image as an output, and

the predetermined image size has a correct answer rate associated with the learned model that falls within an allowable range.

11. A steel manufacturing method comprising:
determining a pass/fail of the steel using the method according to claim 10 ; and
manufacturing steel by controlling a manufacturing condition based on the determination result.

12. The method according to claim 10 , further comprising:
calculating a defect mixing rate of the determination image based on the defect portion; and
determining a grade of the steel corresponding to the determination image, based on the calculated defect mixing rate and a determination result of pass/fail for the steel.

13. The method according to claim 12 , further including a steel manufacturing sorting step of sorting steel for each grade.

14. A non-transitory computer readable medium storing a program that, when executed, causes a computer to perform processing including:
acquiring imaged data for which a surface of steel has been captured;
creating a defect map that is an image indicating a distribution of a defect portion based on the imaged data;
creating a determination image for each defect map by converting an image size of the defect map into a predetermined image size; and
inputting the determination image into a learned model, outputting a value concerning presence/absence of periodic defects, and determining the presence/absence of periodic defects for the determination image based on the value concerning the presence/absence of periodic defects, wherein
the learned model is trained using
a teacher image as an input, the teacher image (i) having the predetermined image size and (ii) being a defect map indicating a distribution of defect portions on the surface of the steel, and
a value concerning presence or absence of periodic d

Assignees

• Jfe Steel Corp

Inventors

• Koshihara Takahiro
• Ono Hiroaki

Classifications

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G01N21/892

  characterised by the flaw, defect or object feature examined · CPC title

• G06N20/00

  Machine learning · CPC title

• G01N2021/8924

  Dents; Relief flaws · CPC title