Confidence-weighted aggregation of spectral priors for RGB illuminant estimation

Color constancy aims to recover the true colors of a scene under unknown lighting conditions, but estimating the light source remains challenging in complex scenes with uneven spatial illumination. This paper addresses image-level illuminant estimation from multispectral images through confidence-weighted aggregation of spectral priors. First, four illuminant spectra (Global Average, White-Patch, Bright Region, and principal component analysis) are computed per-patch from the scene’s multispectral cube and shape-normalized; we show that these statistically distinct priors provide largely overlapping spectral information at the 8×8 patch scale, although their accuracy remains different, while the inclusion of spectral information remains the dominant performance driver under the evaluated NUS-MS protocol. Second, a small-parameter dual-branch network maps the RGB patch and the fused spectral priors to a common 31-dimensional representation and regresses a patch-level illuminant. Finally, a confidence-weighted clustering module on unit-normalized estimates aggregates the patch-level estimates into an image-level illuminant, using patch brightness as the primary reliability signal. Experiments on the 24 chart-masked held-out testing scenes of the NUS-MS dataset show that the proposed method achieves a mean recovery error of 1.14 deg (median 0.18 deg), comparing favorably with classical statistical algorithms and representative learning-based methods under this protocol.