Neural gray edge: Improving gray edge algorithm using neural network

Color constancy is the ability to compute color constant descriptors of objects independent of the light illuminating the scene. Gray-Edge is a recent and color constancy algorithm that is based on this assumption “the average edge difference in a scene is achromatic”. The approximation error of Gray edge increases sometimes because Gray-Edge assumption is not satisfied completely. Therefore, by modeling Gray-Edge assumption, we can compensate the error of Gray-Edge algorithm. In this paper, we proposed a method that is called Neural Gray Edge. This method employs a neural network to model the Gray-Edge assumption based on image statistics. In other words, Gray-Edge acts as a global search that finds the neighborhoods of the scene illuminant vector and then, the neural network acts as a local search and compensates the Gray-Edge error. Experiments on a large dataset of 11000 images show that proposed approach outperforms current state of the art algorithms.