Triangle-based Defect Detection in Perforated Leather using YOLOv5

This paper proposes a novel method based on improved YOLOv5 and triangle area variation for detecting collapsed defects in perforated leather. Defect detection is crucial for ensuring product quality in leather production. The motivation behind this study is to address the issue of collapse in perforated leather and introduce a new detection approach. To enhance the accuracy of detecting small holes, the ASFF feature fusion method is employed to mitigate the adverse effects of PANet's multi-scale feature fusion, along with the incorporation of the CBAM attention mechanism. The improved YOLOv5 is employed to locate the positions of ventilation holes, followed by assessing collapse occurrence through changes in the triangle area formed by adjacent ventilation holes. This method combines the advantages of object detection and area variation analysis, effectively detecting various leather defects, including scratches, wrinkles, and stains, with minor enhancements. Compared to YOLOv3-spp, Faster R-CNN, and YOLOv5, the improved YOLOv5 demonstrates average mAP increases of 4%, 3.7%, and 1.1% for the three defect types, respectively. Moreover, compared to YOLOv5, there is a 1.8% improvement in mAP for small hole detection, and the accuracy of collapse recognition is enhanced by 14.3% with the utilization of the triangle area method. The algorithm's single-image inference time is 62.9ms. Experimental results showcase that this method introduces a novel perspective and approach to the field of leather defect detection, carrying significant practical application value.