Frequency Domain Mask Guided Diffusion Model for Hyperspectral Anomaly Detection

Hyperspectral anomaly detection (HAD) identifies targets with unique spectral characteristics without prior knowledge. In recent years, deep learning-driven HAD methods have pushed the boundaries of innovation, yet these models often falter in accurately modeling the background in complex scenes with noise interference, leading to a challenging conflation of anomalies with the background. Moreover, a significant proportion of deep detection models are finely tuned for specific background contexts, thereby exhibiting a constrained ability to generalize in real-world scenarios. To address these challenges, this study proposes a frequency domain mask-guided diffusion model (FMDiff). By analyzing amplitude spectrum disparities between anomalies and backgrounds, FMDiff constructs a frequency mask that weights background regions (with values close to 1) and suppresses anomalies (with values close to 0). This mask directs a dual-phase diffusion process: in the forward diffusion, it imposes noise selectively on background pixels; in the reverse phase, serving as a condition for the denoising network to effectively estimate and remove background noise. Experimental results illustrate that FMDiff achieves superior performance and is highly competitive with other sophisticated state-of-the-art methods.