ChromaTokenNet : Log‐Chromaticity Token Transformer With Confidence Pooling for Robust Camera Color Constancy

Camera color constancy estimates the chromaticity of the dominant scene illuminant in sensor RGB space so that illumination‐induced color casts can be corrected. Global chromatic statistics are sensitive to scene composition, while unconstrained regressors may exploit camera or dataset‐specific correlations. ChromaTokenNet addresses these limitations by processing aligned semantic and log‐chromatic patch tokens together with differentiable global log‐chromaticity histogram tokens in a shared Transformer. An illumination‐token regressor and a confidence‐pooled patch regressor produce separate log‐chromatic estimates, which are combined by a learned scalar gate. The main contribution is the way these established chromatic, attention, and reliability‐weighting principles are organized in a single model. No outer‐test result generated under that finalized protocol was used to revise the architecture or the training procedure. ChromaTokenNet achieves Mean/Best25/Median/Worst25 recovery angular errors of on ColorChecker RECommended and on INTEL–TAU. Controlled ablations, camera‐held‐out tests, perturbations, and token interventions are used to examine the contribution and limitations of the model components. Performance remains stable under intensity reduction and moderate downsampling, but declines under severe combined low‐light attenuation and additive noise, spatially varying illumination, and non‐invertible clipping.