GPU Color Quantization

There are several algorithms to create a useful color palette, including trainingbased methods such as Self-Organizing Maps (SOM) [Kohonen 01, Chang et al. 05], Linde-Buzo-Gray (LBG) [Verevka 95, Linde et al. 80], and the octree method [Clark 96]. In terms of mean square distortion, the training based methods usually give a much better quantization result. However, training based methods can also be time consuming. While the mean square distortion of LBG and SOM are similar [Leung and Chan 97], the SOM-generated pallet has a very interesting ordering property. To obtain the ordering property, we impose a neighborhood structure among the palette entries before training. After training, when two palette entries are neighbors of each other, they will have a similar color.