A fast codebook design algorithm for vector quantization

This paper presents a fast codebook design algorithm for vector quantization, which eliminates distortion calculations by using inequality between the distortion and the means of two vectors as well as the partial distortion theorem during both the initial codebook splitting and LBG iteration procedures. Moreover, the partition information of vectors in the previous iteration is used in the repartition of the training set to make the employing of the two above eliminating rules more efficiently. This algorithm eliminates about 90% of unnecessary distortion calculations and can produce the same codebook as the LBG algorithm with only a little additional memory, thus outperforming other fast algorithms that always need much more additional memory which is impractical for big codebook cases.