Hyperspectral Image Super-Resolution via Wavelet Decomposition Guidance and Cross-Dimensional Feature Interaction

Hyperspectral images aggregate abundant spectral information, but their spatial resolution is often limited due to constraints inherent in sensor design. Existing super-resolution methods struggle with insufficient high-frequency detail extraction and are restricted by the narrow receptive fields of conventional convolutional operations. To address these issues, we propose a multi-domain reconstruction network guided by wavelet decomposition and incorporating cross-dimensional feature interaction. The network first applies the discrete wavelet transform to decompose the input image into a low-frequency approximation component and high-frequency detail components. The low-frequency component is fed into the 2D branch, where it is fused with locally extracted spatial features to enhance the representation of global structures, while the high-frequency components are processed by the 3D branch to preserve fine-grained spatial details. This design enables collaborative optimization of local and global features. To better exploit subtle spatial variations among adjacent bands, we introduce a band group perception attention mechanism that captures fine-grained differences within spectral groups. Additionally, an efficient 3D multi-scale separable convolution module is designed to extract hierarchical spatial-spectral features via multi-path residual connections. Finally, an adaptive total variation regularization strategy is developed, which dynamically constructs weighting maps based on directional gradient averages to preserve edge structures while reducing over-smoothing. Experimental results on four representative datasets demonstrate the effectiveness of the proposed method in reconstructing both spatial texture details and spectral fidelity.