Enhanced initialization method for LBG codebook design algorithm in vector quantization of images

In this paper, a new initialization method is developed for enhancing the LBG codebook design algorithm in image vector quantization. The proposed method first arranges the training set data according to three different characteristics of the training vector, i.e. mean, variance and shape. A sampling method based on the criterion of maximum error reduction is then developed to select the desired number of representative vectors in the sorted training set as the initial codebook for the LBG algorithm. Computer simulations using real images show that the proposed approach outperforms the random guess and the splitting method. With the new approach, a higher quality of boundary preservation and a better local minimum are obtainable through a fewer number of iteration.