Full Convolutional Color Constancy with Adding Pooling

Abstract

Traditional methods of color constancy have solved the problem by modeling the statistical laws of natural objects and illumination colors. With the development of convolutional neural networks (CNN), the improvements of color constancy have greatly arisen. At first, the CNN-based color constancy algorithms use the images by image patches. However, the illumination information and object information contained in the local image patches may not be enough to estimate the scene illumination color. The full convolutional neural networks (FCNs) architecture can use the entire image as input without thinking about the size if images, which in turn improves the training and testing quality of network. FCNs also allow for end-to-end training to achieve higher efficiency and accuracy. We improve a color constancy algorithm which is based on full convolutional neural network and a new pooling layer named adding pooling. On standard benchmarks, our network outperforms the previous state-of-the-art algorithms.