A new non-linear white balance method is proposed. Taking XYZ stimulus and RGB values of each color block in Munsell color chart illuminated by D65 light source as training samples. A BP neural network mapping relation is built between digit-image RGB values of standard Munsell color chart and XYZ stimulus. Then Estevens matrix is introduced, which is used in translating XYZ stimulus to cone stimulus. After non-linear calculation, the results are translated back into digit-image RGB, values and an image after white balance is produced. It is indicated that this method gives a correct result much closer to image of objects illuminated by standard light sources. In the case of high purity light sources, a large volatility exists in adjusting channel-gains, while a stable result using the proposed method. Especially when the purity of the light source is too high to make values of some color channel approach 0, the channels gain method is out of action, while a good correct result still can be achieved. Meanwhile, the mapping relation between RGB values of digit-image and XYZ stimulus is built, which is a still character for different lighting condition.