Patch Density Estimation for Anomaly Detection with Deep Pyramid Features

Anomaly detection and localization are critical in modern manufacturing for the quality control of products. A particular challenge is that the collecting and labeling of anomaly examples are usually infeasible before implementation. To tackle the problem, a novel two-stage framework is proposed in this paper to build anomaly estimators with normal data only. Specifically, unsupervised deep representations are learned first by a modified SimSiam where an adaptation for one-class learning is implemented. Then the non-parametric method is adopted to model the distribution of training data on the learned representations as the one-class classifier to detect anomaly. Moreover, we model the distribution with different hierarchy level’s features of the convolutional neural network to achieve both image-level and pixel-level detections. Experiments are conducted on MVTec anomaly detection dataset. Competitive results of 92.6% AUROC score for image-level detection and 95.4% for pixel-level detection are obtained to demonstrate the effectiveness of the proposed method.