An Improved Hybrid CoAtNet Architecture for Color Constancy Estimation

Variations in image color caused by different lighting conditions significantly hinder the ability of computer vision systems to perceive true object colors. To address this issue, this paper proposes a light source estimation method tailored for color constancy tasks, built upon an improved CoAtNet hybrid network architecture. In the convolutional stages, the CBAM attention module is integrated to enhance the modeling of local color regions, while the Transformer stages incorporate KAN modules to capture global color dependencies more effectively. To cope with uncertainty in complex environments, a channel-wise confidence weighting module is introduced at the network’s output stage to dynamically aggregate illumination estimates. Extensive experiments are conducted on both the NUS-8 and reprocessed ColorChecker datasets, using a variety of statistical metrics. The results demonstrate that the proposed method outperforms both traditional algorithms and existing deep learning models in terms of accuracy and robustness in color illumination estimation.