Context Quantization Based on the Normality of the Clusters

Abstract

Context is widely used in image coding system to improve the compression performance. In order to achieve the complex high-level statistical correlation of the source effectively, context can be used to get the conditional probability of the current coded symbol. However, it is proved that high-level context is difficult to achieve the true distribution of the signal, as a result the effect of coding is reduced, which is the so-called model cost problem. Context quantization is an efficient method to deal with this problem. As the context quantification is similar to the general vector quantization problem, context quantization can be achieved by the clustering algorithm under the condition that a suitable distortion measure is defined. Currently, K-means is widely used as a clustering algorithm, but K-means algorithm is affected by the premise that the number of classes and the initial cluster centers are given. However, in fact, the optimal number of clusters is often difficult to be defined precisely, resulting in poor clustering effects. Some cluster validity function indicators to this problem have been shown, the main ones are Davies-Bouldin (DB) index, Dunn index, Between-Within-Proportion (BWP) index. Using these cluster validity index is an effective way to find the appropriate number of clusters K, and thus the best number of clusters is obtained. In this paper, the optimal Context quantizer is designed on the basis of cluster validity of K-means.