EGD-Net: Edge-guided and differential attention network for surface defect detection

Establishing an automated defect detection system is a critical task for industrial production, but the current defect detection system still faces great challenges, especially for defects with blurred edges and weak defects in complex backgrounds. To solve these problems, we propose an edge-guided and differential attention network (EGD-Net), which can highlight the defect areas by strengthening edge information and effectively eliminating background clutter. In the proposed network, the multi-scale features are first extracted. Then, a specially designed edge prediction module is used to extract defect edge information from shallow layers. Three multi-scale feature fusion modules are employed to fuse context information in the deep layers. Following this, the edge fusion module is constructed to complement the edge information and context information for better guiding defect segmentation. Finally, a differential attention module (DAM) is designed to perform top-down attention and produce the final prediction results. The DAM can effectively eliminate clutter in the background area caused by connecting the edge features. In the experiment, we collected a packaging box dataset with complex background patterns from the practical industrial field to verify the proposed model performance. Moreover, four public datasets were also employed to validate the model. Experimental results (mIoU/mPA) (DAGM2007: 85.53%/88.19%, CrackForest: 87.58%/91.32%, AITEX defect: 78.31%/82.29%, MT defect: 77.08%/81.19%, box defect 94.39%/96.51%:) show that our proposed method outperforms other state-of-the-art methods, especially for the detection of complex background defects.