Color constancy is the ability of computing color constant descriptors of objects independent of the light illuminating the scene. Although many algorithms exist for color constancy, they are all based on specific assumptions and none of these algorithms can be considered as universal. Gray World and Gray Edge are some samples of such methods. The Gray-World assumption and its variants form a class of color constancy algorithms that are called Gray algorithms. Although the Gray algorithms provide a good approximation of scene illuminant in many times, their performances are low when the Gray assumptions are not satisfied completely. In this article, we proposed a method to improve the Gray algorithms. This method which is called Neural Gray employs a neural network to model the Gray assumptions based on image statistics. In other words, Gray algorithms act as a global search to find the neighborhoods of the scene illuminant vector then the neural network acts as a local search and compensates the algorithms error. Experiments on some benchmark datasets showed that the proposed approach improved the performance of Gray algorithms and it also outperforms state of the art methods. (C) 2013 Wiley Periodicals, Inc.