Enhancing leather defect detection with cross-stage fused attention partial convolution network

Leather surface defects are in various colours, shapes, and sizes. Leather products enterprises usually require low-cost edge computing and embedded devices. These factors pose significant challenges to machine vision-based leather surface defect detection. Under limited computing resources, it is hoped that the learning ability of visual models is enhanced to maintain sufficient accuracy while being lightweight. To this end, the Cross Stage Fused Attention Partial Convolution Network (CSFAPCNet) is proposed which integrates multiple lightweight attention mechanisms to improve feature learning while combining cross-stage partial connections and partial convolutions to reduce computational complexity. Research was conducted on leather defect detection from three levels: leather anomaly detection, multi-type defect detection, and simultaneous detection of positioning and classification. Ablation experiments and generalization performance analysis were also performed. Systematic and in-depth experiments have shown that CSFAPCNet outperforms state-of-the-art methods in various evaluation metrics, and maintains sufficient accuracy while significantly reducing computational complexity.