Work

SSG-GAN: spatial-spectral guided generative adversarial network for hyperspectral reconstruction from RGB images

Year: 2026
Type: article
Abstract: Hyperspectral reconstruction from a single RGB image provides a practical and low-cost solution for spectral imaging with conventional cameras. However, the task is highly ill-posed because three broad RGB channels must be mapped to multiple continuous narrow spectral bands. Existing methods, including convolutional, Transformer-based, multi-scale, and adversarial reconstruction models, have improved reconstruction accuracy, but many of them still learn RGB spatial structures and hyperspectral reconstruction features in a coupled reconstruction stream, while adversarial constraints mainly focus on spatial distributions. To address these issues, we propose SSG-GAN, a spatially guided hyperspectral reconstruction framework. The generator explicitly decouples RGB spatial-prior extraction from hyperspectral reconstruction through a dedicated multi-scale guidance branch. At corresponding decoding stages, a Cross-Branch Spatial-Spectral Attention (CB-SSA) module uses guidance-derived spatial weights to modulate reconstruction features and further refines them through feature-channel modulation and channel recalibration. Moreover, the discriminator incorporates a dedicated pixel-wise spectral-vector branch to evaluate joint band responses, providing complementary adversarial supervision alongside the multi-scale spatial branches. Experiments on CAVE, Harvard, and ICVL demonstrate the competitive performance of SSG-GAN. Specifically, it achieves MRAE/SAM/RMSE values of 0.0605/0.0734/0.0177 on CAVE, 0.0393/0.0294/0.0041 on Harvard, and 0.0303/0.0309/0.00101 on ICVL.
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7234378308
FWCI:
Cites: 57
Cited by:
Related to: 10