A real-time surface defect detection system for industrial products with long-tailed distribution

Abstract

Applying visual recognition algorithms in surface defect detection has aroused increasing interest in industries. Despite the compelling speed advantages over manual detection, many algorithms fail to inspect defects from tail classes, especially where one defect dominates while the others have a few instances. One reason is that most of those computer vision models are proposed for class-balanced datasets while surface defects on industrial products often follow long-tail distributions. Existing studies alleviate this problem by simply adding synthetic data to the tail classes or manually adjusting weights. Herein, we propose: 1) a transformer embedded backbone structure to extract more representative features from the targets; 2) a 3-grids coordinate loss for predicting targets with multi-scale to reduce the targets miss rate. Our system can detect different kinds of surface defects at 125FPS, achieve 9.8% higher mAP and 3-22% higher AP of tail classes than YOLOv4 on long-tailed magnetic tiles datasets. Besides, our experiment on steel plates dataset shows that the effectiveness of our system is not limited to a certain industrial scenario, making it useful for a wide range of automated inspection tasks.