SANFlow: Semantic-Aware Normalizing Flow for Anomaly Detection

Visual anomaly detection, the task of detecting abnormal characteristics in images, is challenging due to the rarity and unpredictability of anomalies. In order to reliably model the distribution of normality and detect anomalies, a few works have attempted to exploit the density estimation ability of normalizing flow (NF). However, previous NF-based methods forcibly transform the distribution of all features into a single distribution (e.g., unit normal distribution), even when the features can have locally distinct semantic information and thus follow different distributions. We claim that forcibly learning to transform such diverse distributions to a single distribution with a single network will cause the learning difficulty, thereby limiting the capacity of a network to discriminate between normal and abnormal data. As such, we propose to transform the distribution of features at each location of a given input image to different distributions. Specifically, we train NF to map the feature distributions of normal data to different distributions at each location in the given image. Furthermore, to enhance the discriminability, we also train NF to map the distribution of abnormal data to a distribution significantly different from that of normal data. The experimental results highlight the efficacy of the proposed framework in improving the density modeling and thus anomaly detection performance.