Diffusing Background Dictionary for Hyperspectral Anomaly Detection

Yaochen Wu, Yu Meng, Lei Sun; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 1046-1064

The diffusion model (DM) has achieved remarkable results in image generation and has been used in hyperspectral image (HSI) processing. However, DM has not been directly applied in the HSI anomaly detection (HAD) task. In this paper, based on the characteristics of HSI and HAD tasks, we combine the advantages of model-driven and data-driven and propose the diffusion background dictionary method (DBD). DBD intrinsically combines the DM with the low-rank representation (LRR) model, using DM to get the crucial background dictionary tensor in the tensor LRR, so that it can accurately detect the anomalies. We also diffuse the multivariate normal distribution that approximates the HSI background based on the idea of the RX algorithm in the HAD, making it more suitable for suppressing the background. DBD combines the advantages of the three main groups of HAD methods, and the experimental results on real datasets prove its effectiveness. DBD can outperform several existing state-of-the-art methods in terms of detection accuracy, which proves the DM's potential in HAD.