High-Resolution Image Anomaly Detection via Spatiotemporal Consistency Incorporated Knowledge Distillation

This paper introduces a new, practical, and challenging scenario of high-resolution (HR) image anomaly detection. Anomaly detection methods cooperated with sliding windows are typical solutions for this task, but they fail to capture long-term dependencies. This paper proposes a Spatiotemporal Consistency Incorporated Knowledge Distillation (STCIKD) method, which translates HR images into video sequences and exploits spatial and temporal consistency between them to capture both local spatial and long-term dependencies. STCIKD consists of a teacher network and two student networks. Among the two students, a spatial student network captures spatial consistency by reconstructing the current video frame, and another temporal student network learns temporal consistency by predicting the future frame. This paper benchmarks several state-of-the-art image anomaly detection methods and evaluates STCIKD for HR image anomaly detection. The results show that by incorporating spatial and temporal consistency, STCIKD significantly outperforms other methods.