Answers are grounded in this paper and cite their sources.
Hyperspectral images often suffer from various degradations during acquisition and transmission. Restoring high-quality hyperspectral images is therefore important for reliable downstream analysis. However, existing methods still struggle to preserve fine structures, especially under complex degradations in remote sensing scenarios. This article proposes a Local-to-Pixel Restoration Network (LPRN) for hyperspectral image restoration. LPRN is a lightweight framework that restores each pixel using its local context. To enhance local feature extraction, we introduce a Wave Mamba module. It combines a distance-aware tokenization strategy and a phase-modulated Wave State Space Model for modeling complex spatial–spectral dependencies. Specifically, local information is first projected into a low-dimensional spectral representation and then encoded into tokens according to distance-aware spatial grouping. The resulting tokens are processed by cascaded Wave Mamba modules for progressive local feature extraction and pixel restoration. Experiments across various restoration tasks, including denoising, noisy super-resolution, and noisy inpainting, show that LPRN achieves competitive performance and efficiency compared with state-of-the-art methods. Visual results further demonstrate its effectiveness in fine-structure restoration.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: