Spectral Super-Resolution From RGB via Dual-Branch Feature Modeling

Abstract

Recovering hyperspectral signals from RGB images is a severely ill-posed problem: reconstruction must preserve coherent material spectra over long spatial ranges while resolving texture-dependent variations near local structures and boundaries. Existing architectures commonly encode both regularities within a shared representation, leaving their distinct spatial requirements only implicitly coordinated. We propose SpecRINet, a dual-branch architecture in which spectral consistency flow propagates long-range material information with linear token complexity, multiscale representative region interaction captures local texture–spectrum correlations, and context-aware gating fuses the two representations according to image content. Across CAVE, Harvard, three scene-disjoint GF5A tests, and one direct cross-sensor GF5B test without fine-tuning, SpecRINet achieves the highest PSNR among the compared methods; its mean margins over the strongest competitor are 0.20 dB on CAVE, 0.63 dB on Harvard, and 0.52 dB on a held-out GF5A scene. Complementary metrics reveal scene- and sensor-dependent tradeoffs, while a boundary-conditioned diagnostic identifies residual error near material transitions.