Steel Surface Defect Segmentation Using U-Net with SLMA-Gray Attention and AdamW Optimization

Abstract

Rust detection on metal surfaces is a challenging task in industrial maintenance and quality control. This paper proposes an improved U-Net segmentation method that enhances the accuracy of surface defect detection in steel materials. This paper proposes an improved U-Net model for steel surface defect segmentation. We introduce a novel Spatial-Channel Gray-Level Mixed Attention mechanism (SLMA-Gray) to enhance defect saliency in grayscale images and employ the AdamW optimizer to improve generalization. Experiments on the NEU-DET dataset show our method significantly outperforms the original U-Net, with increases of 5.29% in precision, 18.84% in recall, 3.34% in accuracy, 15.17% in Dice, and 18.18% in IoU.