Аннотация
Hyperspectral image reconstruction from RGB inputs is a compelling yet complex task due to high spectral dimensionality, scarcity of labeled data, and limitations of traditional models in capturing both spectral and spatial dependencies. Existing approaches often exhibit poor domain generalization, high computational cost, and degraded perceptual quality, particularly when transferring across scenes with different spectral distributions. HS-RankFormer is introduced as a lightweight transformer-based framework for robust RGB-to-HSI reconstruction. The framework incorporates three key components: a rank-aware transformer generator that encourages low-rank spectral structure while learning global spectral–spatial correlations; an edge-guided fusion strategy that augments RGB inputs with Sobel-derived edge maps to enhance structural awareness; and a self-supervised masked pretraining scheme that enables effective representation learning without paired annotations. Together, these components improve spectral accuracy, spatial consistency, and training efficiency. Performance is evaluated on four benchmark datasets: ARAD 1 K, BGU-HS, CAVE, and KAUST-Skin. Comparative experiments against representative state-of-the-art methods, including MST++, AWAN, Restormer, and pix2pix, show competitive or superior performance across MRAE, RMSE, PSNR, SSIM, SAM, and ΔE2000 metrics. Qualitative results further illustrate improved perceptual fidelity and strong cross-domain robustness. By combining transformer-based modeling with spectral regularization and edge-enhanced learning, HS-RankFormer offers a robust and generalizable approach for RGB-to-HSI reconstruction, with consistent performance across remote sensing, biomedical imaging, and conservation monitoring datasets.