A robust weakly supervised learning of deep Conv-Nets for surface defect inspection

Automatic defect detection is a challenging task owing to the complex textured background with non-uniform intensity distribution, weak differences between defects and background, diversity of defect types, and high cost of annotated samples. In order to solve these challenges, this paper proposes a novel end-to-end defect classification and segmentation framework based on weakly supervised learning of a convolutional neural network (CNN) with attention architecture. Firstly, a novel end-to-end CNN architecture integrating the robust classifier and spatial attention module is proposed to enhance defect feature representation ability, which significantly improves the classification accuracy. Secondly, a new spatial attention class activation map (SA-CAM) is proposed to improve segmentation adaptability by generating more accurate heatmap. Moreover, for different surface texture, SA-CAM can significantly suppress the background's inference and highlight defect area. Finally, the proposed weakly supervised learning framework is trained using only global image labels and devoted to two main visual recognition tasks: defect samples classification and area segmentation. At the same time, it is robust to complex backgrounds. Results of the experiments verify the generalization of the proposed method on three distinct datasets with different kinds of textures and backgrounds. In the classification tasks, the proposed method improves accuracy by 0.66-25.50%. In the segmentation tasks, the proposed method improves accuracy by 5.49-7.07%.