GND-YOLOv7 Based Leather Defect Detection Algorithm Research

In the leather manufacturing industry, online detection of leather defects can significantly reduce production waste. This thesis proposes a lightweight algorithm for detecting leather defects based on GDN (Ghost-DFC Attention-NWD) to improve YOLOv7 to solve the problems of lacking defection samples and improving detection speed and accuracy. First, the data volume was enhanced using the Albumentations and Mosaic. Then, to reduce the parameters scale, the Ghost module is adopted to replace the Convolution module in Backbone and Neck. Finally, to improve the detection accuracy of small targets, the DFC Attention mechanism is used to enhance the feature extraction capability, while the NWD standard is used to replace the commonly used IOU metric. The experimental results show that the proposed algorithm decreases the amount of computation by 44.36% and the number of parameters by 37.81 % compared to YOLOv7, and achieves 90% AP for the small target defects and 87.8 % overall mAP.