A self-looping semi-supervised surface defect detection method for steel parts

By employing a self-looping training and dynamic data synthesis method, and utilizing the TransUNet and Transformer architecture, the problems of sample scarcity and noise accumulation in steel component defect detection are solved, achieving high-precision detection of minute defects.

CN122391087APending Publication Date: 2026-07-14CHINA TOBACCO ZHEJIANG IND CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
CHINA TOBACCO ZHEJIANG IND CO LTD
Filing Date
2026-03-27
Publication Date
2026-07-14

AI Technical Summary

Technical Problem

Defect detection of steel components faces the problems of scarce defect samples and accumulation of false label noise, making it difficult for traditional detection methods to achieve high-precision identification of minute defects.

Method used

We employ a hybrid architecture combining TransUNet segmentation network with Transformer and CNN. Through self-loop training and dynamic data synthesis, we construct a clean background library and a defect sample library using a small amount of labeled data. We use a hybrid loss function and exponential moving average strategy to remove noise and dynamically expand the defect sample library.

Benefits of technology

It significantly improves the accuracy of defect detection in steel components with a small amount of label data, increases the detection rate of minor defects, reduces the impact of false label noise, and approaches the level of full supervision inspection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122391087A_ABST
    Figure CN122391087A_ABST
Patent Text Reader

Abstract

The application designs a self-circulation semi-supervised surface defect detection method for steel parts. In view of the problems of defect sample scarcity and pseudo-label noise accumulation of steel parts, the application firstly constructs a pure background library by using image restoration technology. In the training stage, a mathematical synthesis model based on "Copy-Paste" is established, and synthetic training data is dynamically generated through random geometric transformation, scale scaling, illumination disturbance and Gaussian noise injection. At the same time, the unlabeled data is inferred by using the teacher model updated by the exponential moving average, and the high-confidence pseudo-defects are screened based on the multiple threshold values of confidence, area and circumscribed rectangle geometric features, and the defect library is dynamically expanded. The application combines the global modeling capability of TransUNet and the mixed loss function optimization, and effectively suppresses the noise interference through multiple rounds of self-circulation iterative algorithm. Experiments show that under the condition of only 10% labeled data, the average intersection over union is improved to 0.7845, which is significantly better than the traditional semi-supervised method.
Need to check novelty before this filing date? Find Prior Art