基于深度学习的带钢表面缺陷检测

打开文本图片集

关键词：带钢表面缺陷检测；深度学习；编码器－解码器结构；残差细化网络；Dropout正则化；工业质量控制中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2025）19-0172-05

Abstract：Aiming at the issues of low eficiencyand insufficient accuracy in traditional surface defect detection methodsforstel strips，thisresearchproposesaDeepLearningmodel integratinganencoder-decoderstructure，Residual Refinement Network，and Dropoutregularization.The encoder-decoder architecture enances feature extractionand defect localizationcapabilities，whiletheResidual RefinementNetworkmitigates gradientvanishingissuesandimprovesdeep feature representation.And Dropoutregularzationsuppresss overfitingandenhances modelgeneralization.Experimentsconductedon the NEU-DET and Xsteel datasets demonstrate that the proposed method achieves an accuracy of 95% in the test set， with recall and precision rates of 93% and 97% ，respectively， improving by 13% and 12% compared to the baseline model， and an F1-score of 95% .Thisresearch significantly enhances the precision and robustnessof defect detection forsteelstrips，providing technical support for industrial quality control.

Keywords： surface defect detection for stee strips; Deep Learning; encoder-decoder structure; Residual Refinement Network;Dropout regularization; industrial quality control

0 引言

随着工业4.0的浪潮以前所未有的速度席卷全球，智能制造与质量控制已经跃升到了钢铁工业生产中最为关键且不可或缺的核心地位。（剩余7773字）