Redução de complexidade do processo de decisão de modo da predição intra-quadro do codificador de vídeo VVC utilizando aprendizado de máquina

Abstract

This work presents a solution for reducing the complexity of the intra mode decision for the Versatile Video Coding (VVC). To achieve this complexity reduction, the order in which modes are evaluated in the Rate-Distortion Optimization (RDO) process has been changed in order to give priority to intra prediction modes that are more likely to be chosen as the best at the end of the decision process. By doing this, it was possible to insert supervised machine learning models based on binary decision trees to predict when it is possible to skip the evaluation of some types of intra modes based. A total of three decision tree-based models were developed. The first and second models were developed with the objective of predicting when Angular intra prediction modes or Matrix-based Intra Prediction modes (MIPs) can be skip from the RDO evaluation without introducing significant losses in coding efficiency. The third model was developed to act in cases where the first and second models were not able to skip the evaluation of any types of modes, seeking to predict if MIP modes should be evaluated or not. For the first, second and third models, accuracies of 85.27%, 78.69% and 84.80% were obtained. As a result, we obtained averages of 10.87% on complexity reduction and 0.39% of BD-BR increase. When compared with related works, it can be noticed that even though the developed method has lower complexity reductions in some cases, it obtains lower coding efficiency losses in all cases, even when considering all intra modes available in the VVC.