Hyperspectral Image Super-Resolution via Grouped Second-Order Spatial Features and Spectral Attention Network

Hyperspectral imaging (HSI) captures detailed spectral data across a wide range of wavelengths, making HSI invaluable for many computer vision applications. However, limitations of imaging systems result in low spatial resolution in HSI images. State-of-the-art research encounters challenges due to complex textural patterns and inadequate low-frequency details in captured images. Moreover, high-fidelity image reconstruction is computationally intensive. To address these issues, we developed a grouped second-order spatial features and spectral attention network (GSSAN). GSSAN incorporates a grouped second-order spatial attention (GSOSA) module to extract relevant first- and second-order spatial features. We employed a grouped spectral attention (GSA) module to extract spectral details across different bands of HSI images. Rotary positional embedding (RoPE) is incorporated into our model to encode relative distances and angular rotations between features while preserving equivariance. These modules help reconstruct clear edges and contours in HSI. Our experimental results demonstrate that GSSAN significantly outperforms state-of-the-art HSISR methods, achieving improvements in mean peak signal-to-noise ratio (MPSNR) by up to 0.183 dB and mean structural similarity index (MSSIM) by 0.051 on the Chikusei test dataset with a scaling factor of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\times 4$</tex-math></inline-formula> . The average inference time of our model is 16.09s. These findings indicate an improved trade-off between reconstruction quality and computational efficiency.