Color Constancy Based on Multi-Scale Multi-Task Convolutional Neural Networks

Color constancy aims at recovering true surface color from a captured image. The key procedure is to estimate light color and correct for any color bias that it introduces. Over the past few decades, numerous methods for illumination estimation have been introduced. Although they have gained great progress, feature extraction is still a bottleneck for accurate illumination estimation. Inspired by the deep convolutional neural network (CNN) paradigm that can extract those suitable features for the task on the fly, rather than predefine a set of features, we present a new multi-scale multitask CNN (MM-CNN) for illumination estimation. Different from those typical CNN solutions, our proposed model estimation can jointly supervise scene classification and illumination in a united framework. In addition, a multi-task training scheme with deep supervision is also introduced to enhance the training efficiency and accuracy of the MM-CNN model. Experimental results on two real-world image datasets demonstrate that the MM-CNN-based method can significantly enhance color constancy performance.