A Hybrid Model of State-Space Model and Attention for Hyperspectral Image Denoising

Hyperspectral images (HSIs) exhibit pronounced spatial similarity and spectral correlation. With these two physical properties taken into account, underlying clean HSI will be easier to derive from noisy images. However, existing denoising approaches struggle to model the spatial-spectral structure due to the following limitations: excessive memory consumption when performing global modeling, and insufficient effectiveness in local modeling. To address these issues, in this article, we propose HyMatt, a hybrid model of the state-space model and attention mechanism for HSI denoising. Specifically, to fully exploit global similarity within an HSI cube, we devise vision Mamba quad directions based on crafted cube selective scan (CSS) to capture long-range dependencies in a memory-efficient manner. Our CSS not only enhances global modeling capacity but also mitigates the negative impacts of causal modeling inherent in the SSM. Furthermore, in order to improve local similarity modeling, we integrate a local attention module, in which the adjacent elements are refined by adaptively utilizing similar neighboring features as guidance. Compared to existing methods, our HyMatt excels in exploiting local features while leveraging the global similarity within the entire HSI cube. Extensive experiments on both simulated and real remote sensing noisy images demonstrate that our HyMatt consistently surpasses the state-of-the-art HSIs denoising methods.