A Defect Detection Method for Artificial Leather Based on YOLOv5

In recent years, the demand for high-quality artificial leather has increased, highlighting the need for accurate and efficient defect detection methods. To address issues of inefficiency and a high leakage rate m manually inspecting defects in artificial leather, this study introduces a defect detection approach using an enhanced YOLOv5 network. In this method, the Ghost convolution module replaces the CNN convolution module within the CSP structure of the backbone network. This adjustment simplifies the network architecture, enhances model feature extraction comprehensiveness, and reduces the parameter count in the backbone network. Additionally, the SE attention mechanism is incorporated into the backbone network, allowing it to learn the texture and edge features of small target leather defects more effectively and rapidly, thereby improving the network’s sensitivity to minor defects Finally, by substituting the original CIoU loss function with the EIoU loss function, the method intensifies the penalty on discrepancies between the predicted and actual bounding boxes, specifically in width and height. Experimental results indicate that the modified model achieves an overall average accuracy (mAP) of 87.1% on the tested dataset, offering a more efficient solution for detecting defects m artificial leather.