Redução do Tempo de Codificação na Predição Interquadros do Padrão Versatile Video Coding (VVC)

Abstract

There is a growing demand for high-resolution digital videos, mainly over the In ternet. Different platforms allow the transmission, sharing, and consumption of videos online. During the COVID-19 pandemic, this context intensified, demonstrating the importance of network infrastructure and strategies for video compression. In this sense, the Versatile Video Coding (VVC) standard provides new tools that enable higher compression ratios than previous standards and greater adaptability to different types of videos. The VVC standard and other modern encoders are based on the hybrid model of prediction residual coding. Its main steps include intra-frame and inter-frame prediction, Transforms, Quantization, and Entropy coding. However, along with these characteristics, VVC is highly complex, resulting in a very high video encoding time. Therefore, many challenges are still open to reduce this problem. In light of this, this thesis aims to present solutions for reducing the encoding time in inter-frame prediction of the VVC standard. In addition to providing context on the VVC standard and inter-frame prediction itself, concepts of machine learning are also presented, along with results from a dense experimental analysis and a systematic literature review. Finally, four optimization solutions focused on inter-frame prediction are presented: one heuristic based on statistical analysis and three other solutions employing machine learning. The heuristic focuses on unidirectional, bidirectional, and Affine predictions, and can be configured for three operating points. This so lution achieved encoding time reductions of up to 26.5% for the unidirectional and bidirectional stages and up to 22.71% for the Affine stage, with a coding efficiency loss of 0.9% and 0.44%, respectively. The solution using Random Forest focused on bidirectional achieved an average reduction of 92% in time per step and 2% in total encoding time, with a coding efficiency loss of 0.75%. The expanded solution, adding Decision Tree models focused on Affine to the previous solution, achieved a total time reduction of 3.9%, with a coding efficiency loss of 0.72%. Finally, the solution focused on Affine optimization, using Decision Tree models, obtained an average reduction of 42.1% in time per step, 3.4% in total time, with an impact of 0.25% on coding efficiency. Therefore, these results demonstrate that the use of machine learning techniques are the most promising ways to achieve a reduction in interframe coordination time with minimal impacts on update efficiency.