A fast codebook generation algorithm for vector quantization of images

A fast codebook generation algorithm on a 2-D DCT (discrete cosine transform) is presented for vector quantization of images. A 2-D DCT is used to extract the feature of training images. Using the good energy compact capability of the DCT, certain significant components of the feature space are selected to partition the training vectors into a binary tree. For a nonterminal node of the binary tree, a single feature with the largest variance and its corresponding mean value are adopted as the split feature and the split threshold, respectively. Simulation results show that the new algorithm requires about 51% of the amount of time needed by the pairwise nearest neighbor algorithm, and a better reconstructed picture quality is obtained