MTSR: Mamba-Transformer Super-Resolution Model for Hyperspectral Images

Deep learning-based methods, particularly Transformer-based approaches, have made remarkable achievements in hyperspectral image super-resolution (SR) tasks. However, due to the complex terrestrial object structures and rich spatial-spectral details in hyperspectral images (HSI), solely using the Transformer-based backbone with large scale parameters makes it challenging to effectively represent such diverse features and often leads to substantial computational resource consumption. To address these issues, this article proposes an efficient Mamba-Transformer super-resolution network (MTSR) for single HSI, which combines state-space modeling with the multiheads self-attention mechanism to highlight both the efficiency and accuracy of SR restoration. Specifically, this scheme designs a multiscale dilated attention block that integrates features across multiple scales by attention mechanisms, enabling the capture of fine distinctive spatial-spectral features. Moreover, the Mamba-Transformer block (MTB) sequentially incorporates Mamba modules with Transformer modules to facilitate effective extraction of spatial and spectral features. In this structure, MTB refines spatial features with two Mamba modules while a Transformer module captures global spectral information by exploiting spectral correlations. Extensive experiments demonstrate that MTSR approach significantly improves the SR performance in comparison with the baseline and achieves a favorable tradeoff between SR restoration accuracy and computational complexity.