RGB-Guided Multi-Kernel Attention Feature Extraction Network for Hyperspectral Image Super-Resolution

Hyperspectral image (HSI) super-resolution aims to reconstruct high-spatial-resolution images from their low-resolution counterparts while preserving critical spectral fidelity. Existing single-modal methods inevitably encounter information bottlenecks, often struggling to recover high-frequency spatial details. To overcome this, multi-modal approaches integrate auxiliary RGB images; however, most existing paradigms heavily couple the modalities. This excessive coupling not only creates a strong dependency on the auxiliary modality but also allows structural discrepancies and noise to propagate into the reconstructed HSI, causing severe spectral distortion. To address these limitations, this paper presents RGB-GMKAFENet (RGB-Guided Multi-Kernel Attention Feature Extraction Network for Hyperspectral Image Super-Resolution), a novel framework that treats the auxiliary RGB image purely as a structural guide rather than a heavily coupled fusion counterpart. The proposed architecture introduces three key innovations: (1) a Multi-Kernel Attention Feature Extraction (MKAFE) module incorporating cosine similarity (CSKA), multi-scale spectral gradient (MSSGKA), and spectral correlation coefficient (SCCKA) kernel attention for comprehensive spectral-spatialfeature modeling with linear computational complexity; (2) an RGB-Guided Feature Injection Module (RGB-GFIM) that utilizes global patch indexing and cross-modal cross-attention to selectively inject high-frequency RGB texture details into HSI features; and (3) a VGG perceptual loss function that constrains feature-space differences in the RGB domain to enhance visual quality. Extensive experiments on the CAVE, Pavia Center, and Pavia University datasets demonstrate that RGB-GMKAFENet yields highly competitive and robust performance across multiple evaluation metrics (PSNR, MPSNR, SSIM, SAM) at×2, ×4, and×8upscaling factors. The results confirm that our guided strategy effectively improves spatial detail recovery while strictly maintaining the intrinsic spectral fidelity of the HSI.