A two-stage network based on edge information for visual anomaly detection
Di Chen, Juan Xu · 2023
Reconstruction-based approaches attract increasing attention in unsupervised visual anomaly detection. Such methods rely on the assumption that reconstruction models trained using only normal images are not well reconstructed for anomalous data, so that the anomalous regions can be identified by computing the differences in the reconstructions. However, it's usually challenging to manage the model's generalization boundaries in practice. The model with a poor generalization capability will produce additional image differences in the normal regions, while the model with an overly strong generalization capability can even well reconstruct the abnormal regions, making the abnormal regions less distinguishable from the normal regions. In this paper, we propose a two-stage reconstruction network to solve the above problems, which can generate high-accuracy and anomaly-free reconstruction. Specifically, our two-stage anomaly detection network consists of the edge extraction sub-network and the image restoration sub-network. The first sub-network generates anomaly-free grayscale edges of the image as intermediate results, and then the second one recovers normal color and low-frequency information for the intermediate results. The proposed network can generate high-accuracy and anomaly-free reconstructions to maximize the difference between the reconstruction errors of normal and abnormal samples, which can ultimately improve the accuracy of anomaly detection. Experiments on the dataset MVTec AD show that our proposed method achieves competitive results in anomaly detection and anomaly localization (97.8% for detection and 97.7% for localization, AUROC).