Wavelet-Enhanced Weighted and Dense Cross-Scope Transformer for Hyperspectral Image Super-Resolution

Hyperspectral image super-resolution is crucial to enhance the spatial detail of hyperspectral data and has achieved notable advancements in these years. The recently state-of-the-art cross-scope spatial-spectral Transformer(CST) method for single-image super-resolution can capture long-range spatial-spectral dependencies through cross-scope self-attention mechanisms effectively. However, CST lacks effective feature interaction across layers, limiting its ability to fully exploit hierarchical spatial-spectral information. In order to overcome this shortage, this paper proposes a novel Wavelet-Enhanced Weighted and Dense Cross-Scope Transformer (WWDCST) for single-image super-resolution. This approach combines the strengths of weighted connection and dense cross-scope transformer layer (Dense-CSTL) in capturing features across different layers with the feature extraction and denoising capabilities of wavelet transform. WWDCST is designed to extract the hierarchical features from each convolutional layer adaptively, while the importance of individual layers is dynamically adjusted through self-adaptive parameter learning. Considering the denoising and feature extraction advantage of wavelet transform, it is integrated into the spatial domain processing pipeline to improve the super-resolution accuracy of hyperspectral images. Comprehensive experiments show that our proposed network architecture surpasses existing algorithms in both quantitative metrics and visual perception.