Color quantization is a process of sampling three-dimensional color space (e.g. RGB) to reduce the number of colors in a color image. By reducing to a discrete subset of colors known as a color codebook or palette, each pixel in the original image is mapped to an entry according to these palette colors. In this paper, a reinforcement-learning approach to color image quantization is proposed. Fuzzy rules, which are used to select appropriate parameters for the adaptive clustering algorithm applied to color quantization, are built through reinforcement learning. By comparing this new method with the original adaptive clustering algorithm on 30 color images, our method shows an improvement of 3.3% to 5.8% in peak signal to noise ratio (PSNR) values on average and results in savings of about 10% in computation time. Moreover, we demonstrate that reinforcement learning is an efficacious as well as efficient way to provide a solution of the learning problem where there is a lack of knowledge regarding the input-output relationship.
Tamkang Journal of Science and Engineering=淡江理工學刊 14(2), pp.141-150