Visual Differential-Spatially Projected Transformer for Efficient Hyperspectral Images Super-Resolution

Single hyperspectral image super-resolution aims to restore high-resolution hyperspectral images (HSIs) from a single low-resolution HSI. Existing methods often struggle with inefficient spatial attention mechanisms and inadequate modeling of cross-spectral dependencies, which hinder the simultaneous achievement of high restoration accuracy and computational efficiency. To address these limitations, we propose an efficient yet powerful architecture named Visual Differential-Spatially Projected Transformer (VDSPT). Instead of conventional spatial attention, VDSPT introduces Visual Differential Spatial Attention (VDSA), which computes two distinct attention maps and subtracts one from the other to suppress overemphasized regions. This differential strategy enables more balanced and effective extraction of spatial features. Additionally, we design Spatially Projected Spectral Attention (SPSA), which projects features along the spatial dimension before computing spectral attention while preserving the natural ordering of spectral channels. This leads to more robust and reliable modeling of inter-channel spectral relationships. By stacking only a few VDSA and SPSA modules, VDSPT achieves the state-of-the-art super-resolution performance with remarkable computational efficiency. Extensive experiments on both natural and remote sensing hyperspectral datasets demonstrate that VDSPT consistently outperforms existing methods in terms of quantitative metrics and visual quality. The source code will be publicly released.