MMRR: Unsupervised Anomaly Detection through Multi-Level Masking and Restoration with Refinement

Jaesung Ahn, Janghyeon Lee, Hanbyel Cho, Yooshin Cho, HyeongGwon Hong, Junmo Kim

Keywords: Deep Learning, Computer Vision, Anomaly Detection

Abstract: Recent state-of-the-art anomaly detection algorithms mainly adopt generative models or approaches based on deep one-class classification. These approaches have hyperparameters to balance the adversarial framework of the generative adversarial network and to determine the decision boundary of the classifier. Both methods show good performance, but their performance suffers from hyperparameter sensitivity. A new category of anomaly detection methods has been proposed that utilizes prior knowledge about abnormal data or pretrained features, but it is more generic not to use such side information. In this study, we propose "Multi-Level Masking and Restoration with Refinement (MMRR)", an unsupervised-learning-based anomaly detection method based on a generative model that overcomes hyperparameter sensitivity and the need for side information. MMRR learns the salient features of normal data distributions through restoration from restricted information via masking, resulting in a better restoration of in-distribution data than out-of-distribution data. To overcome hyperparameter sensitivity, we ensemble restoration results from information restricted to predefined multiple levels instead of finding a single optimal restriction level, and propose a novel mask generation and refinement method to achieve hyperparameter robustness. Extensive experimental evaluation on common benchmarks (i.e. MNIST, FMNIST, CIFAR10, MVTecAD) demonstrates the efficacy of the MMRR.

TL;DR: We propose a novel Multi-Level Masking and Restoration with Refinement (MMRR) to solve the hyperparameter sensitivity problem overlooked by existing anomaly detection studies.