Visible anomaly detection, an essential drawback in pc imaginative and prescient, is often formulated as a one-class classification and segmentation activity. The scholar-teacher (S-T) framework has proved to be efficient in fixing this problem. Nevertheless, earlier works based mostly on S-T solely empirically utilized constraints on regular knowledge and fused multi-level info. On this research, we suggest an improved mannequin referred to as DeSTSeg, which integrates a pre-trained trainer community, a denoising pupil encoder-decoder, and a segmentation community into one framework. First, to strengthen the constraints on anomalous knowledge, we introduce a denoising process that enables the scholar community to be taught extra strong representations. From synthetically corrupted regular photos, we prepare the scholar community to match the trainer community function of the identical photos with out corruption. Second, to fuse the multi-level S-T options adaptively, we prepare a segmentation community with wealthy supervision from artificial anomaly masks, reaching a considerable efficiency enchancment. Experiments on the commercial inspection benchmark dataset reveal that our technique achieves state-of-the-art efficiency, 98.6% on image-level ROC, 75.8% on pixel-level common precision, and 76.4% on instance-level common precision.
