Method for detecting surface defect, method for training model, apparatus, device, and media

Patent metadata

Field	Value
Publication number	US-11763552-B2
Application number	US-202017116597-A
Country	US
Kind code	B2
Filing date	Dec 9, 2020
Priority date	Jun 12, 2020
Publication date	Sep 19, 2023
Grant date	Sep 19, 2023

Abstract

A method for detecting a surface defect, a method for training model, an apparatus, a device, and a medium, are provided. The method includes: inputting a surface image of the article for detection into a defect detection model to perform a defect detection, and acquiring a defect detection result output by the defect detection model; inputting a surface image of a defective article determined to be defective into an image discrimination model based on the defect detection result to determine whether the surface image of the defective article is defective, wherein the image discrimination model is a trained generative adversarial networks model, and the generative adversarial networks model is obtained by training using a surface image of a defect-free good article; and adjusting the defect detection result of the surface image of the defective article according to a determination result of the image discrimination model.

First claim

What is claimed is:

1. A method for detecting a surface defect of an article, the method comprising:
inputting a surface image of the article for detection into a defect detection model to perform a defect detection, and acquiring a defect detection result output by the defect detection model;
inputting a surface image of a defective article determined to be defective into an image discrimination model based on the defect detection result to determine whether the surface image of the defective article is defective, wherein the image discrimination model is a trained generative adversarial networks model, and the trained generative adversarial networks model is obtained by training using a surface image of a defect-free good article; and
adjusting the defect detection result of the surface image of the defective article according to a determination result of the image discrimination model.

2. The method according to claim 1 , wherein the defect detection result output by the defect detection model comprises: a coordinate of an external rectangular frame of a defect comprised in an article surface, a coordinate of a contour of the defect, and a category of the defect.

3. The method according to claim 1 , wherein the adjusting the defect detection result of the surface image of the defective article according to a determination result of the image discrimination model comprises:
confirming the defect detection result output by the defect detection model in response to the discrimination result determining that a defect exists in the surface image of the defective article; and
rejecting the defect detection result output by the defect detection model in response to the discrimination result determining that no defect exists in the surface image of the defective article.

4. The method according to claim 1 , wherein before the inputting a surface image of the article for detection into the defect detection model to perform the defect detection, the method further comprises:
in response to determining that a resolution of the surface image of the article for detection is inconsistent with an input image resolution of the defect detection model, segmenting the surface image of the article for detection to match the input image resolution.

5. The method according to claim 1 , wherein the article is a machined part.

6. A method of training an article surface defect recognition model, the method comprising:
acquiring a surface image of a defect-free good article as a discriminating training sample; and
inputting the discriminating training sample into a generative adversarial networks model for training, wherein the generative adversarial networks model comprises a generative networks model and a discriminating network model, wherein the generative networks model is configured for generating a new surface image of the good article based on the discriminating training sample, and the discriminating network model is configured for discriminating based on the generated surface image of the good article; and the generative networks model is configured for determining whether a defect exists or not in a surface image of a defective article with a detected defect at a detection stage.

7. The method according to claim 6 , further comprising:
acquiring a surface image of a good article corresponding to at least one type of article surface respectively, and acquiring at least one type of defect image, wherein the defect image is an image of a defect on a surface of the article;
combining the good article surface image and the defect image to generate a defect article surface image as a defect training sample; and
inputting the defect training sample, and a good article surface image used as a good article training sample into a defect detection model for training, wherein the defect detection model is configured for detecting a defect in the surface image of the article.

8. The method according to claim 7 , wherein the combining the good article surface image and the defect image to generate the defect article surface image comprises:
combining the good article surface image and the defect image to generate the defect article surface image based on a non-conditional generation model.

9. The method according to claim 7 , wherein the acquiring a surface image of a good article corresponding to at least one type of article surface respectively, and acquiring at least one type of defect image comprises:
shooting, using a shooting apparatus, at least one surface of at least one good article and at least one defective article respectively to acquire a surface image of the good article and a surface image of the defective article; and
extracting a partial image comprising the defect from the defective article surface image as the defect image, and labeling a defect type of the defect image.

10. The method according to claim 9 , wherein the combining the good article surface image and the defect image to generate the defect article surface image comprises:
combining a plurality of surface images of the good article with a plurality of types of defect images respectively to generate a plurality of surface images of the defective article, wherein each surface image of the defective article is one type of article surface comprising one type of defect.

11. The method according to claim 7 , wherein before the inputting the defect training sample, and a good article surface image used as a good article training sample into a defect detection model for training, the method further comprises:
segmenting the defect training sample and the good training sample according to an input image resolution of the defect detection model to retain partial images with a same resolution as the input image resolution, and updating the partial images as the defect training sample and the good training sample.

12. The method according to claim 6 , wherein the article is a machined part.

13. An electronic device comprising:
at least one processor; and
a memory in communication with the at least one processor;
wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform operations, the operations comprising:
inputting a surface image of an article for detection into a defect detection model to perform a defect detection, and acquiring a defect detection result output by the defect detection model;
inputting a surface image of a defective article determined to be defective into an image discrimination model based on the defect detection result to determine whether the surface image of the defective article is defective, wherein the image discrimination model is a trained generative adversarial networks model, and the trained generative adversarial networks model is obtained by training using a surface image of a defect-free good article; and
adjusting the defect detection result of the surface image of the defective article according to a determination result of the image discrimination model.

14. The electronic device according to claim 13 , wherein the defect detection result output by the defect detection model comprises: a coordinate of an external rectangular frame of a defect comprised in an article surface, a coordinate of a contour of the defect, and a category of the defect.

15. The electronic device according to claim 13 , wherein the adjusting the defect detection result of the surface image of the defective article according to a determination result of the image discrimination model comprises:
confirming the defect detection result output by the defect detection model in

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Lin Shufei
• Zhu Jianfeng
• Yuan Pengcheng
• Zhang Bin
• Han Shumin
• Xu Yingbo
• Feng Yuan
• Xin Ying
• Wang Xiaodi
• Liu Jingwei
• Wen Shilei
• Zhang Hongwu
• Ding Errui

Classifications

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0475

  Generative networks · CPC title

• G06N3/094

  Adversarial learning · CPC title

• G01N21/8851 Primary

  Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges (G01N21/8806 and G01N21/93 - G01N21/95692 take precedence; optical measurement of dimensions G01B11/00; optical scanning G02B26/10; image transformation G06T3/00; computerised image enhancement G06T5/00; image processing per se for flaw detection G06T7/0002) · CPC title