Unsupervised Spectral Super-Resolution Guided by Spectral Sampling Priors

Spectral super-resolution (SSR) has garnered significant attention in recent years. Most existing networks rely on supervised methods, which require paired RGB and hyperspectral images (HSIs) for training. However, HSI acquisition is costly and time-consuming due to specialized hardware and complex preprocessing. Additionally, spectral mixing phenomena in low-resolution HSIs degrade image quality. To address these challenges, spectral super-resolution (SSR) techniques have emerged to generate high-quality HSIs from widely accessible RGB images, enabling applications in agriculture, medicine, and environmental monitoring. To address these issues, we propose a novel Unsupervised Spectral Super-Resolution Network guided by Spectral Sampling Priors (<bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">SPointNet</b>). Inspired by multi-modality text-image fusion techniques, we first introduce the Point-Image Fusion module (PI-Fusion), which fuses sampled spectral data with RGB images. We then utilize Spectral Unmixing for Super-Resolution module (UMSR) to produce a coarse HSI, maximizing the exploitation of spectral information. Finally, we integrate a Multistage Shuffle-Unshuffle Transformer (MSUT) to fuse the coarse HSI with the RGB image, enhancing its spatial information. SPointNet can ensure continuity and consistency in both spectral and spatial dimensions in the generation of the refined HSI, which is validated on three publicly available datasets.