Light-Ad: Learning a Lightweight Detector for On-Board Hyperspectral Anomaly Detection via Target-Background Decoupled Knowledge Distillation

On-board hyperspectral remote sensing anomaly detection can identify surface anomalies on edge platforms without prior, serving important application areas which have high requirements for the transferability and real-time performance. However, transferable anomaly detectors have large parameter and high computational cost, resulting in the contradiction between the transferability and efficiency. To deal with this problem, we designed a lightweight hyperspectral anomaly detection method based on target-background decoupling distillation for real-time processing on edge. Specifically, we employed a knowledge distillation framework, in which the student network is designed as an anomaly-specific sub-network and a normal-specific sub-network, to learn the knowledge of target and background, respectively. Additionally, a target-background deviation loss is proposed to further enable the network to learn the differences between targets and backgrounds. Experiments conducted on public datasets demonstrate that our lightweight network reduce the number of parameter by 13 times with minimal loss in accuracy, leading in both precision and processing efficiency compared to existing hyperspectral anomaly detection methods and knowledge distillation methods.