Context-based K-means And Mixed Ant Colony Clustering Algorithm Quantitative Research

In recent years, Context-based entropy coders are included in many image coding algorithms. According to the conclusion that conditional entropy will be no more than the unconditional entropy, we can see the Context model can effectively reduce the entropy of the source of information, thereby reducing the code length of the image coding. But facts proved that a too large Context model will increase the difficulty of counting statistics of the source symbol, as a result, coding efficiency will be reduced. This is the problem which is often encountered. That is so-called "Context dilution" in the Context model entropy. Therefore, a critical issue in context-based image coding systems is how to resolve the conflict between the desire for large templates to model high-order statistic dependency among pixels and the problem of context dilution because of insufficient sample statistics for a given input image.In order to solve this problem, quantization of the context models will be a good option. In order to make the conditional probability distribution of Context model in the Context Quantization easier to statistics and better convergence in the true probability distribution of source of information, it needs to classify the conditions distribution of the Context model following certain rules which have been established. The classification method is commonly clustering algorithm. Most partition-based clustering algorithms including the classic K-means clustering algorithm are subject to the local optimum problem, so how to make a clustering algorithm to find the global optimum becomes the focus of many researchers.In this paper, we present an improved context quantization algorithm based on a hybrid K-means and ant colony clustering algorithm. The K-means clustering algorithm is used to construct the initial solution for a context quantization problem. An ant colony based on clustering algorithm is used to improve the quality of the solution. During each iteration, objects are assigned to respective clusters based on the corresponding pheromone concentrations updated by the artificial ants. Then, a local search procedure is conducted by a small part of the ants with the best objective function values to further refine the solution. Experiment results show that the presented algorithm outperforms context quantization algorithm based on the K-means clustering and the Maximum Mutual Information under various quantization levels.