Patch-based dual-channel weighted image color constancy network

Due to the limitations of unknown light sources, object surface reflectivity, and imaging devices, it is still challenging to infer scene illumination based on the color representation of object surfaces. To further optimize the deep learning-based color constancy algorithm, we propose a two-channel weighted image color constancy network based on image patches. The network utilizes a feature fusion technique to combine shallow edge features and deep fine-grained features extracted from two network branches respectively. The strategy of using image patches increases the diversity of the dataset, and the weighting mechanism suppresses image noise to further improve the performance of the network. Experimental results on reprocessed Color Checker and NUS-8 datasets show that the network significantly improves performance while maintaining model complexity.