RGB color constancy from multispectral images via hierarchical physics-injection and spectral augmentation

Computational color constancy (CCC) aims to recover the canonical appearance of scenes, yet current spectral-based approaches are hindered by the scarcity of hyperspectral data. This limitation compels reliance on suboptimal two-stage strategies, restricting the potential of deep learning. To overcome this, we propose a physically accurate spectral data augmentation strategy that expands the training set five-fold using real-world illuminants. Leveraging this, we introduce the Hierarchical Physics-Injection Network (HPI-Net). Experiments demonstrate that, within the full-multispectral-input setting, HPI-Net achieves a 20% reduction in mean recovery angular error relative to the strongest compared baseline, validating the effectiveness of our end-to-end physics-injection architecture.