AutoAno: Anomaly Localization with Self-supervised Multi-scale Feature and Multivariate Gaussian Estimation
Qiang Tong, Meixue Sun, Bo Wang, Dianyu Liu · IEEE/WIC/ACM International Conference on Web Intelligence · 2021
Anomaly localization in images has long been challenging and valuable for both research and real-world applications. Popular methods for anomaly localization are with Auto-Encoder variants and existing GAN-based models, and have obtained decent outcomes. However, these generative models have significant limitations, such as the over-powerful generalization capacity in anomaly data. To address this problem, in this paper, we propose AutoAno, a self-supervised learning model that leverages both multi-scale feature and multivariate Gaussian estimation to achieve robust anomaly localization. In detail, our method extracts multi-scale features from the patches of an image, which can simultaneously consider global, contextual, and local features, so that contains richer semantic information to anomaly localization. We then employ multidimensional Gaussian distribution to estimate the low-dimensional representation for the normal samples in training data. Notably, we discard the commonly used reconstruction loss, and instead use the distance between multi-scale feature and the estimated Gaussian distribution to detect and localize the anomalies. To verify the effectiveness of our work, we compare AutoAno with the prevailing models with extensive experiments on benchmark and real-world datasets. The empirical results demonstrate that our model outperforms the state-of-the-art models significantly.