Multi-scale Texture Network for Industrial Surface Defect Detection

Automated surface defect detection is crucial for ensuring product quality in industrial settings. This paper presents a multi-scale texture network that addresses this challenge by effectively analyzing textures at various scales. The proposed network incorporates a “Multi-Scale Texture Feature Processing” module to generate multi-scale texture tokens for comprehensive surface analysis. Additionally, a “Multi-Head Feature Encoding” mechanism captures local details and global patterns, leading to improved accuracy. Besides, we introduce a multi-scale perceptual loss function that guides training by optimizing images at different scales while preserving perceptual similarity. Experimental results demonstrate the effectiveness of our approach, offering high accuracy in automated surface defect detection.