Mixed Context Techniques in the Adaptive Arithmetic Coding Process for DC Term and Lossless Image Encoding

Context modeling plays a very important role in data compression. In the arithmetic coding process, one usually classifies the data to be encoded into several contexts according to the characteristic of the causal part. Each context has a frequency table, which reflects the probability distribution of the data belonging to this context. However, in practice, it often happens that the characteristic of the causal part is between that of two contexts. For example, if we assign the context according to whether the average gradient of the causal part is greater or smaller than 10, when the average gradient is around 10, if we assign the case into one of the contexts, since the input is not typical for both two contexts, the coding gain may not be achieved. In this work, we proposed an idea of mixed context. That is, in the intermediate case, the frequency table is constructed from that of the two adjacent contexts. Moreover, after encoding the data, the frequency table of both the two contexts are adjusted. Experiments for the example of JPEG DC term and lossless image encoding show that, with the proposed algorithm, the coding efficiency can be much improved and the frequency table converges to the truth probability distribution quickly.