Robust weakly supervised product surface defect segmentation based on guided cropping and inpainting extension

In the realm of industrial manufacturing, automatic detection and segmentation of surface defects in products is vital to enhance both product quality and efficiency. However, a large number of existing deep learning methods require a substantial amount of manually labeled data for training, and the high-cost labeling process hampers the practical application of such methods. Towards this end, we present a weakly supervised defect segmentation algorithm without any segment labels. First, a training enhancement method based on a contrastive learning module (CLM) coupled with a guided cropping module (GCM) is proposed to improve the network's attention to defects in the Defect Focus Classifier (DFC) training phase. Subsequently, a novel inpainting extension module (IEM) generates a final class activation map (CAM) to obtain a pseudo label automatically for segment network training. Finally, conditional random field (CRF) and an additional training round refine the segmentation results. Moreover, the whole process is distilled into a fully supervised segmentation network to improve the inference efficiency. Conducting extensive experimentation, we have achieved 100% and 93.39% average precision (AP) and 41.73% and 53.14% average intersection-over-union (IOU) on the public datasets KolektorSDD and KolektorSDD2, respectively. Furthermore, we verified the generalizability of our method by conducting experiments on several industrial product classes in the MVTec AD, MTD, and DAGM datasets. In these experiments, we achieved favorable classification and segmentation results using solely image classification labels.