Single Hyperspectral Image Super-Resolution Using Multi-scale Cross-spectral Transformer Network

Abstract

Hyperspectral image super-resolution seeks to produce images with finer spatial detail while keeping spectral information consistent, which supports more reliable analysis in hyperspectral imaging applications. However, existing methods mainly rely on 2D/3D convolutions, which are limited to modeling local spatial-spectral representations and fail to fully capture global spectral dependencies across bands. To address this limitation, we develop a Multi-scale Cross-spectral Transformer Network, named MCTNet. For local representation learning, MCTNet introduces a position-guided multi-scale attention fusion module. This module combines spatial feature extraction and channel attention. It extracts contextual information at multiple scales and highlights informative spectral changes. For global representation learning, CSTM applies cross-spectral self-attention to model dependencies between distant latent spectral-dimension feature channels. Experiments on standard benchmark datasets demonstrate that MCTNet provides reconstruction accuracy and visual quality that are comparable to or better than those of recent state-of-the-art approaches.