Abstract
As the experiments presented in Chap. 4 have shown and the complexity scaling methods of Chap. 5s have confirmed, an important share of the high computational complexity of HEVC comes from the use of very flexible partitioning structures, such as the CUs, the PUs and the TUs. This chapter describes the process of using data mining (DM) techniques to build a set of models that are used to decide if the RDO-based partitioning structure decision process should be terminated early or run to its full extent [1–4]. By using information from intermediate encoding variables collected during the encoding of a set of video sequences, a set of decision trees were built and implemented in the HM encoder. When using this modified encoder, the operation of the decision trees sidesteps the encoder from having to run the full RDO process to find the best partitioning structures. The study of correlations and information gains associated with each variable, recorded while encoding test videos with the original HM encoder, was essential to the development of the early termination schemes presented in this chapter.