Our eyes are very good at removing the effect of illumination to judge the true color of an object. A simple way of modeling the effect of a light is to assume that the reflected color I = (ir; ig; ib) arises due to a light L = (lr; lg; lb) interacting with paint C = (cr; cg; cb) satisfying L * C = I, i.e. at each pixel we have (lr x cr; lg x cg; lb x cb) = (ir; ig; ib). Computing L and C given I is ill-posed. For example, a red pixel in an image I = (255; 0; 0) might be due to a white light L = (1; 1; 1) on a red object C = (255; 0; 0), or red light L = (1; 0; 0) on a white object C = (255; 255; 255). Thus, in order to solve this problem certain priors are needed. One such prior is the \gray world" assumption that states the average color of the image under white light is gray (Recall that any color (r; g; b) where r = g = b is gray).
Assuming that the average color of an image under white light L = (1; 1; 1) is (128; 128; 128). Under this assumption, given an image the color of the light can be computed as L = (rave/128 ; gave/128 ; bave/128 , where, rave; gave; bave are the average red, green, and blue values of the image.
Thus,the true color of a pixel can be obtained as: cr = ir x 128/rave; cg = ig x 128/gave; cb = ib x 128/bave; WhiteBalance.m takes an image I and returns the light L and color image C using the above calculations.