In this article we propose an original technique to reduce the number of colors contained in an image. This method uses the “Baker's Transform” which allows to obtain a statistically suitable mixture of the pixels of the image. From this mixture, we can extract a square sample, which constitutes a good representation of the initial image. From a sample or from all the possible samples, we can obtain the pallet allowing an adaptive reduction of the number of the colors. We consider four effective methods to obtain this pallet. Finally, we obtain results, which illustrate the good visual quality reached by the reduced images.