Unsupervised defect inspection algorithm based on cascaded GAN with edge repair feature fusion

Surface defect inspection can greatly improve the efficiency of industrial production by replacing manual operations. However, in actual industrial scenarios, it is difficult to collect and manually label enough defect images. In addition, the complex backgrounds, diverse shapes and sizes, and broad random location distribution of defects in images make defect inspection more challenging. To address these issues, we propose an unsupervised defect inspection algorithm based on cascaded GAN (Generative Adversarial Networks) with edge repair feature fusion. In this algorithm, the edge repair network provides intact structural features for the defect repair network by means of a feature fusion method based on channel attention. For the edge repair network, we develop a deformable autoencoder, which fully utilizes the ability of deformable convolution to perceive very little contextual information to improve its ability to repair defect edges. Specifically, training requires only a few defect-free images and no labeled defect images. To verify the effectiveness of the proposed algorithm, we compare it with existing algorithms in terms of precision, the F1-measure, and the mIoU (mean Intersection over Union). The experimental results show that the proposed algorithm exhibits better defect inspection performance, especially for defects with rich forms and diverse positions against complex backgrounds.