Spectral reflectance imaging with dual-illumination and RGB camera via regularized end-to-end learning

Spectral reflectance imaging reveals illumination-independent material properties, supporting accurate analysis in areas like medicine and cultural preservation. However, traditional hyperspectral systems are costly, and single-RGB methods often struggle with spectral ambiguity and limited accuracy. To address this trade-off between accuracy and efficiency, we propose a dual-illumination-based spectral reflectance imaging method using RGB inputs. Our method introduces a regularization-guided end-to-end framework that jointly optimizes illumination selection and reflectance reconstruction. Specifically, we incorporate regularization terms derived from multi-illumination gain priors to guide discriminative illumination learning, and design a physically-aware illumination modeling network to alleviate optimization imbalance between branches. For reconstruction, we build a lightweight yet high-performance architecture that integrates adaptive illumination augmentation, hierarchical multi-path processing, and channel-wise recalibration. Complementary spectral and spatial loss functions further improve reconstruction quality. Experiments show that our method significantly outperforms state-of-the-art methods on multiple datasets and enables high-quality reflectance reconstruction even with low-cost hardware. We deploy a physical prototype system and demonstrate its reliability under hardware-induced spectral deviations, as well as its practicality and deployment potential through the downstream application.