Improving RGB-to-hyperspectral reconstruction accuracy using residual frequency-spatial integration network

Hyperspectral reconstruction aims to recover dense spectral information from RGB images, providing a cost-effective alternative to direct hyperspectral image acquisition. Although existing convolutional neural network (CNN)-based methods have achieved promising performance, most of them mainly rely on spatial-domain feature extraction and local convolutional operations, which limits their ability to capture long-range spatial dependencies and complex inter-band spectral correlations. To address these limitations, we propose a Residual Frequency-Spatial Integration Network (RFSI-Net) for RGB-to-hyperspectral image reconstruction. The key contribution of RFSI-Net lies in introducing a frequency-aware feature modulation strategy into this specific reconstruction task. The proposed frequency modulation module transforms spatial-spectral features into the frequency domain through a 3D Fourier transform and adaptively modulates the frequency-domain representations. This provides a frequency-domain representation complementary to spatial-domain convolution, aiming to improve the modelling of global spatial structures and inter-band spectral correlations. In addition, we design a dual-attention module that combines channel attention and spatial attention to further enhance discriminative feature representation, allowing the network to focus on informative spectral channels and key spatial regions. A frequency-domain consistency loss is also introduced to constrain the reconstruction process from the frequency perspective, thereby improving spectral correlation modelling and reconstruction fidelity. Extensive experiments on four benchmark datasets demonstrate that the proposed RFSI-Net outperforms existing state-of-the-art methods on multiple evaluation metrics. In particular, on the remote sensing dataset GDFC, our method achieves a 5.1% improvement in MRAE and a 10.7% reduction in SAM, demonstrating its effectiveness in both spectral accuracy and spatial detail preservation.