Visual abnormality detection, a major problem in computer vision, is usually formulated as a class segmentation and classification task. The Student-Teacher (ST) framework has proven effective in solving this challenge. However, previous ST-based work only applied empirical constraints on normal data and multilevel fused information. In this study, we propose an improved model called DeSTSeg, which integrates a network of pre-trained teachers, a denoising student encoder, and a segmentation network into a single frame. First, to strengthen the constraints on outliers, we introduce a denoising procedure that allows the student network to learn stronger representations. From synthetically damaged normal images, we trained the student network to match the teacher network characteristic of the same images without corruption. Second, to merge the functions of multilevel ST adaptively, we trained a segmentation network with rich supervision from synthetic anomaly masks, achieving substantial performance improvement. Experiments on the industry inspection reference dataset show that our method achieves state-of-the-art performance, 98.6% ROC at the image level, 75.8% average accuracy at the pixel level, and 76.4% in average precision at the instance level.
