Um esquema rápido baseado em aprendizado de máquina para a predição interquadros do codificador de vídeo VVC

Abstract

The new trends in digital video technologies, such as 4K-8K resolution, omnidirectional, and screen content, have become popular in video coding scenarios. To support these technologies, new video encoders have been proposed in recent years. The Versatile Video Coding (VVC) is the most recent video coding standard proposed by the Joint Video Experts Team. Released in July of 2020, the new standard achieves a coding efficiency of 33% higher when compared to its predecessor, the HEVC. This efficiency is achieved by implementing new coding tools, such as flexible partitioning tools, new prediction modes for intra-blocks, and new transform modes. However, the implementation of new tools comes at the cost of a significant encoding time increase. Also, to improve the coding efficiency, VVC implements the affine motion estimation, which allows the detection of non-translational transformations during the interframe prediction. However, the affine motion estimation increases the complexity of VVC interframe prediction up to 47% and becomes an obstacle for real-time applications. In this sense, this work proposes a scheme to reduce the complexity of the VVC encoder by the flow control of the steps of affine motion estimation. Named as LEAP (Learning-based Affine Prediction), the proposed scheme is based on machine learning, and, when implemented in the VVC reference software, it is capable of reducing the total encoding time by 8.49%, and the encoding time of the affine motion estimation by 46.94%, on average, with minor penalties in coding efficiency. Besides, LEAP can be combined with other techniques to improve the complexity reduction of VVC.