Self-Supervised Surface Defect Detection Methods for Industrial Products

Compared with normal images, abnormal images have less proportion and higher acquisition cost. Thus, most existing methods are based on learning a normal image auto-encoding model. An image is considered to contain abnormal areas if it cannot be reconstructed well. Although this kind of method has made some progress, this framework can only ensure that the model can reconstruct the image as high as possible according to the input image, and cannot guarantee that the model can learn the ability to find abnormal areas in the input image. Naturally, if the neural network can learn abnormal data, the model will have better distinguishability. Addressing it, in this paper, we supplement abnormal data through data enhancement, which is trained together with normal data to realize self-supervised learning. Although the exceptions made by data enhancement do not perfectly simulate the real defects, the model can still learn a good ability to catch exceptions due to the generalization ability of deep learning. Besides, the ECA-Net is embedded into the backbone network which is capable of capturing complementary semantics and improving the accuracy of anomaly detection. In addition, CAM-loss drives the backbone to express the features of the target category, to obtain more discriminative feature representations that can get a better image-level localization by Grad-CAM. Extensive quantitative and qualitative experiments on the large-scale industrial product anomaly detection dataset (i.e., MVTec) demonstrate that the proposed method can achieve state-of-the-art performance.