Residual Flow Group Attention Network For Leather Defect Classification

As the demand for leather increases, automated leather defect identification is becoming increasingly important in the leather industry. Based on the issues that unobvious color of leather defects, small defect area and low recognition efficiency, this paper proposes a residual flow group attention network (ResFGANet) to effectively identify leather defects. First, group convolution and spatially separable convolution are introduced in feature extraction part to compress model parameters and improve the efficiency of defect recognition. Second, an improved efficient attention module is added after feature extraction to improve the recognizability of defect features in a lightweight way. Finally, this paper conduct a series of comparative experiments and ablation experiments on two defective data sets, and the results show that our method can achieve the best results with a small number of parameters.