Abstract: In vector quantization（VQ），the initial codebook design influences or determines the times of iteration of codebook training method and the quality of codebook.So it is very important for VQ codebook.For existing initial codebook algorithms，the quality of initial codebook is strongly influenced by the initial codewords selected from the training vectors and initial codebook is difficult to match with the features of the training vectors.To overcome these disadvantages，an algorithm，which sorts training vectors according to the sum of each training vector components，divides the sorting vectors to some separating area，calculates the mean of the vectors of each area to obtain the initial codewords，is proposed.Because of the use of vector feature，the proposed algorithm doesn’t depend on the structures of images and produces a robust initial codebook.Experimental results show that the proposed algorithm is a better algorithm.The proposed method can be combined with the LBG algorithm to further improve the quality of codebook.

Key words: the sum of vector, initial codebook, vector quantization, image processing

摘要： 矢量量化的初始码书设计是很重要的，影响或决定着其后码书形成算法的迭代次数和最终的码书质量。针对原有的初始码书算法在性能上随机性强与信源匹配程度不高的问题，提出一种对于训练矢量实施基于分量的和值排序，然后做分离平均的初始码书形成算法。算法使用了矢量的特征量，脱离了对于图像结构因数的依赖，能产生鲁棒性较好的初始码书。实验证明了该方法的有效性，与LBG算法结合可进一步提高码书质量。

关键词: 矢量的和值, 初始码书, 矢量量化, 图像处理