Probabilistic Graphical Model Based Fast HEVC Inter Prediction

In this article, we propose a probabilistic graphic model based fast HEVC encoding framework. A Bayesian network is characterized by the structure of the network (nodes and edges in the graph) and the probabilistic distributions. It can be constructed in three main steps: 1) data collection and pre-processing, 2) learning network structure, 3) learning parameters of the probabilistic distributions. In the first step, we select a subset of possible HEVC encoding parameters to be modeled by the Bayesian network. Data were collected using the first 150 frames of the HEVC common test condition Class D sequences. Then the model structure is trained so that it is consistent with the conditional dependencies from the observation. Finally, we use a Gaussian Bayesian network to properly model both discrete and continuous valued variables. Based on the Bayesian network and parameters, we can calculate the conditional probabilities for different status in the encoding process and neglect the events with probabilities smaller than a threshold. In this way, a fast HEVC encoding is achieved.