A Grouping-Based Spectral Reconstruction Network Through Frequency-Domain Attention

To break through the hardware limitations, spectral super-resolution (SSR) is an effective way to reconstruct hyperspectral images (HSIs) from RGB images. However, current methods tend to recover tens of hundreds of bands from the only three channels of RGB images, which ignores the spectral correlations and distinguished high-frequency (HF)/low-frequency (LF) spatial patterns, resulting in inconsiderable reconstruction. In this article, a novel grouping-based spectral reconstruction network (GBSRNet) is proposed, in which the complex entire image reconstruction is separated into a number of subimage reconstructions through adopting a divide-and-conquer strategy. Specifically, for each subreconstruction, a spatial feature reconstruction module (SFRM) is proposed for discriminative learning of HF/LF spatial patterns through integrating a parallel Transformer with the Fourier transform to disentangle the spatial features. Thus, more precise representations of distinguished HF/LF spatial counterparts can be obtained. Furthermore, to improve the spectral fidelity, a spectral feature enhancement module (SFEM) incorporated with local and global learning is designed to efficiently enhance the local self-similarities and global diversities of spectral features. Comprehensive experiments on several datasets illustrate that the newly proposed GBSRNet clearly outperforms some state-of-the-art SSR methods, generating HSIs from an RGB image with better spatial quality and spectral fidelity.