Decisão rápida na predição intra-quadro do codificador AV1 usando aprendizado de máquina

Abstract

The playback and transmission of high-definition digital videos grows exponentially worldwide. These videos are used for both entertainment and professional purposes and have the internet as their main source of propagation. Considering the volume of information needed to represent a video, the massive space to store digital media and the huge data traffic generated, it is essential to use efficient video compression techniques. In this way, video encoders focus on eliminating data redundancies and reducing information that is of little relevance to the human visual system. The use of video encoders standarized by international standardization organizations, generally involve a high cost due to the expenses of royalties applied. In an attempt to avoid these costs, Aomedia was founded. Aomedia is a technology industry consortium that launches in March 2018 the specification of its royalty-free encoder, called Aomedia Video 1 (AV1). However, this developed format achieves high values of coding efficiency at the same time as it increases the computational costs that reflects the time taken to enconding and the related energy consumption. This work presents a solution for the fast mode decision in the intra-frame prediction process of the AV1 encoder, reducing the computational cost involved in decision-making of the predictor mode, using supervised machine learning. Random Forest-based models were trained on datasets created from the experiments performed in the AV1 reference software, libaom. The main goal was to reduce the modes evaluated in intra prediction, accelarating the decision process. Then, it was possible to implemented the models at the encoder and evaluate the impacts in terms of encoding time reduction and encoding efficiency losses. As a result, it was possible to achieve averages of time savings of 44% with and impact of 4.6% on BD-Rate.