This paper studied the methods dealing with the distortions and deviations in color which are caused by the factors such as the limitations on imaging-forming principle, device performance and machining controls This paper compared color restoration by the BP neural network versus the global polynomial regression. Then this paper presented an RGB to L*a*b* transformation method based on polynomial regression of each subspace through dividing the space into sub-domains in accordance with the hue. The calculated average color difference based on BP neural network was 2.8476, and the difference based on global polynomial regression was 2.857; the two had only 0.3% difference. However, after using polynomial regression of each subspace to recover the colors, the average color difference was 2.206, reduced by 23% compared with the above two methods. Recovering the colors by using polynomial regression of each subspace can effectively improve the precision of color restoration.