A Novel Weakly Supervised Learning Method for Industrial Surface Defect Detection

Surface defect detection plays a decisive role in controlling product quality which is an important part of achieving automated industrial production. To solve the problems of the high cost of pixel-wise labeling effort, this work proposes a novel weakly-supervised method for surface defect detection which only requires image-level labels. Our approach only uses image-level labels and a small number of samples for training, which is mainly divided into three folds, (1) image classification, (2) class activation map (CAM) generation or CAM modification, (3) foreground and background segmentation. Firstly, ResNet-38 with similar classification performance and better segmentation performance is selected as the backbone of a modified CAM network. Then, we introduce the self-supervised equivariant attention mechanism (SEAM) method, which contains equivariant regularization and pixel context correlation modules to solve the problem of CAM over-activation and under-activation. Extensive experimental results demonstrate the effectiveness of our method.