Deep Learning‐Based White Balance Correction Using Spectrum Estimation From Multispectral Imaging

ABSTRACT

Maintaining color consistency across different illuminants is a long‐standing problem in digital imaging. In this context, color consistency refers to a system's ability to map a given surface to the same output color regardless of the capture illuminant. Residual color errors can persist under spectrally complex illumination. RGB inputs provide limited spectral information to separate illuminants with similar appearance but different spectra. In this study, we propose a three‐stage white balance correction framework that combines multispectral imaging (MSI) with deep learning. The framework integrates illuminant classification, class‐specific spectral regression, and an imaging pipeline informed by the predicted spectral power distribution (SPD). To address class imbalance, class‐weighted learning is employed during classification, and both shared and class‐specific architectures are evaluated for SPD prediction. The proposed method is compared against two baselines: direct SPD regression without classification and uncorrected RGB images converted to sRGB. Experiments on real‐world data show that the classification‐guided framework significantly reduces color error relative to the spectroradiometric reference, with decreasing from 14.1 to 4.6 under LED lighting and from 4.9 to 1.5 under fluorescent conditions. The model also achieves improved stability (lower variance) and spectral prediction accuracy using class‐specific architectures. While limitations remain, such as dataset imbalance and the lack of value‐level augmentation, these results demonstrate the feasibility of MSI‐based white balance correction and underscore its potential to enhance perceptual color fidelity in consumer imaging.