Dual-Branch Network With Mutual Guidance for Hyperspectral Image Superresolution

Fusion-based hyperspectral image super-resolution has recently attracted increasing interest due to its superior reconstruction quality. This approach enhances the spatial resolution of low-resolution hyperspectral images (LR-HSIs) by fusing high-resolution multispectral images (HR-MSIs) of the same scene. However, most existing deep learning-based methods have not sufficiently considered the huge modality differences between these two types of images before feature fusion, potentially resulting in the loss of valuable information. To tackle this challenge, we introduce a novel dual-branch network with mutual guidance (DBMGNet). Specifically, we employ a dual-branch architecture to process the input LR-HSI and HR-MSI in parallel, with the aim of reconciling their modality differences. For this purpose, we designed a mutually guided dual-stream Transformer block that performs bidirectional calibration between the branches, enhancing their spatial and spectral consistency. To prevent excessive coupling of information, we proposed a multiscale feature enhancement block, which independently refines fine-grained details within each branch. Finally, a weighted feature fusion block is developed to effectively integrate the features from both branches. Experiments on three widely used datasets indicate that the proposed DBMGNet achieves stable and superior performance with lower computational cost in comparison with several state-of-the-art approaches.