An Improved Background Suppression Diffusion Model for Hyperspectral Anomaly Detection

As advances Artificial Intelligence (AI) technology, there has been an increasing focus on utilizing Deep Learning (DL) techniques for Hyperspectral Anomaly Detection (HAD).Anomalies are detected by HAD methods through the analysis of Hyperspectral Image (HSI) data.Recent studies have proposed various methods based on Diffusion Model (DM) to implement background suppression in HAD, particularly by utilizing the reverse inference process of Denoising Diffusion Probabilistic Models (DDPM).Furthermore, these DDPM-based methods often overlook the spatial relationships between pixels, which are essential for achieving accurate anomaly detection results.To address this limitation, this paper introduces two improvements for original background suppression diffusion model.First, we added corresponding weights to the loss function as well as to the output of the network.This allows the optimized model to focus on local spatial features involved in training, thereby reducing the reconstructive effect on the anomalous pixel of the HSI.Secondly, we introduce a nonlinear mapping that integrates the detection results from the diffusion model with the most discriminative band in HSI.This mapping enhances the distinction between the background and anomalies, significantly improving detection accuracy.Experimental results on three public datasets show that our method outperforms other state-of-the-art detectors in overall performance.