Multi-Scale Retinex algorithm is one of the effective traditional algorithms on illumination compensation to face image of poor quality. but the phenomenon such as the white and color distortion, low contrast, which might be produced during process of the retinex algorithm to face images. In order to improve the face recognition rate under complex light conditions, this paper proposes a new and multi-scale retinex with color restoration for light compensation of color images. On the basis of studying of MSRCR theory, bettering the original algorithm, and using the fast Fourier algorithm, so its speed is faster than that of the traditional method. Experiment proved the contrast, brightness, hidden details of images are enhanced by using the improved algorithm. This paper describes the principle and implementation method of the algorithm and at the same time compares it with histogram equalization, Gamma transform methods through experiments. Results of experiments show that good effects can be produced while using the algorithm for light compensation of color images.