Lightweight Detection Model for Animal Wet-Blue Hide Surface Defects Based on Yolov5s

Abstract

In the process of animal leather processing, the surface damage of wet-blue hides restricts the quality of leather products. To ensure the efficiency and quality of animal leather processing, a lightweight model for detecting surface defects on wet-blue hides based on optimized YOLOv5s is proposed. The new model adopts the lightweight EfficientNetV2 network to extract surface defect features and incorporates a spatial pyramid pooling–fast (SPPF) structure at the end of the network to obtain features at different scales. Efficient multi-scale attention (EMA) was embedded in the bottom-up structure of the Neck section to achieve comprehensive feature extraction and retention, ensuring that spatial semantic features are adequately distributed in each feature. A dataset of wet-blue hide defects was constructed and used to verify the performance of the new model. the experimental results show that, the new model is superior to the commonly used classical detection models. The precision rates for detecting three types of leather surface defects, namely imprint, puncture, and breakage, are 86.5%, 95.3%, and 87.9%, respectively. These results can provide technical support for research of surface damage detection in other leather processing applications.