Clustering-based cycle GAN for camera anomaly detection

Abstract

The goal of Camera Anomaly Detection (CAD) in this paper is to detect and locate anomalous areas, so as to assist in judging whether surveillance cameras need to be replaced or repaired. In practice, most automatic CAD methods are implemented based on statistical modeling, frequency-domain analysis, sparse representation, low-rank decomposition, convolutional neural network (CNN) models, and so on. However, as an one-class problem, it is difficult to balance the model’s generalization ability and anomaly discovery capability. To address this dilemma, this paper proposes a novel end-to-end Cluster-based Cyclic GAN (CCGAN). Since anomalous samples are usually scarce in practice, the proposed method follows a fully unsupervised learning strategy that only requires a small number of anomaly-free images for model optimization. It introduces a cycle mechanism to improve the identity mapping ability on anomaly-free regions, and uses high-dimensional clustering to softly discriminate the boundary between encoded normal and anomalous features. Experimental results show that, in terms of precision, recall and F1-measure, the proposed method outperforms comparative approaches in pixel-level anomaly detection.