CGSCNet: ConvGRU Based Skip-connection Framework for Hyperspectral Image Super-Resolution

Hyperspectral image super-resolution has been extensively studied, and significant development has been made based on deep convolutional neural networks. However, in most residua super-resolution networks, straightforward skip-connections spatial features with different levels of abstraction can lead to semantic confusion, ultimately affecting reconstruction accuracy. To address this issue, we developed a ConvGRU based skip-connection strategy. This strategy leverages ConvGRU to filter features across consecutive layers, enhancing feature propagation and fusion. Accordingly, we develop a ConvGRU based skip-connection framework designed specifically for hyperspectral image super-resolution (CGSCNet). The framework consists of three parts. The first module is spatial feature reconstruction, which introduces ConvGRU to handle the fusion of semantic information across different layers and gradually builds spatial-spectral information by leveraging the feature reuse properties of the residual network. The second module is detail refinement, aimed at enhancing spatial detail information. The third module, spectral information reconstruction, recalibrates spectral responses by modeling the relationships between different spectral bands. Experiments on the Chikusei, and Houston datasets demonstrate that our framework outperforms state-of-the-art comparison models in both quantitative evaluation metrics and visual quality across a wide range of scenarios.