Color constancy is a critical aspect of visual perception, enabling consistent color recognition under varying lighting conditions. However, achieving reliable color constancy remains a significant challenge, especially in scenes characterized by complex illuminations or insufficient information to determine a unique or even limited range of illumination colors. These challenges often lead to inaccuracies in color perception, impacting various applications in computer vision, such as object recognition, image processing and visual scene understanding. This paper presents a novel dual-branch model to address the problem of color constancy. The first branch of the model takes an image as input and employs a triplet attention mechanism as a feature extraction network to capture spatial and contextual information. Meanwhile, we calculate the log-chroma histogram of the input images and extract features using the SqueezeNet-based parallel branch, focusing on the distribution of color information. Features from both branches are then fused according to a dual affinity matrix to predict the illumination. Experiments on Reprocessed Color Checker Dataset and NUS-8 Dataset demonstrate that our model achieves superior performance in color difference estimation compared to existing methods, achieving the median of angular errors of \(0.90^\circ \) and \(0.86^\circ \) , along with the Worst 25% of angular errors of \(1.73^\circ \) and \(2.02^\circ \) , which highlight the effectiveness and robustness of our model.