Aprendizado de Máquina Aplicado à Redução do Custo Computacional do Test Zone Search e da Estimação de Movimento Affine do Codificador VVC

• Autor(a) principal: Viana, Ramiro Gomes da Silva
• Data de Publicação: 2025
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFPel - Guaiaca
• Texto Completo: http://guaiaca.ufpel.edu.br/xmlui/handle/prefix/15988
• topic: Predição inter-quadros; Versatile Video Coding; Aprendizado de máquina; Árvores de decisão; Inter frame prediction; Machine learning; Decision trees; CIENCIAS EXATAS E DA TERRA; CIENCIA DA COMPUTACAO

As the demand for video transmission surges on remote work, education, and streaming services, the need for continuous advancements in video encoding tech nologies becomes increasingly evident. Adapting to the evolving requirements of efficient video delivery and consumption necessitates ongoing development and en hancement in video encoding standards, with Versatile Video Coding (VVC) emerging as a notable example. This work provides an overview of the main algorithms in VVC Inter-Frame Prediction, focusing primarily on Test Zone Search (TZS) and Affine Motion Estimation (AME), two of the most computationally intensive tools within VVC. Furthermore, this work introduces an approach to accelerate TZS and AME using Machine Learning, specifically employing Decision Trees. First, an acceleration of the VVC Test Model (VTM) reference software was performed, focusing on TZS, by selectively skipping its last three steps using a set of 12 Decision Trees, one for each block size supported by TZS in VVC. Next, the acceleration of the VTM’s AME was performed, considering the accelerated implementation of TZS. In this case, the entire AME process is selectively skipped using a new set of 12 Decision Trees, one for each block size supported by AME in VVC. The proposed approach achieved an average reduction of 20.99% in the total VVC encoding time, an average reduction of 62.15% in the TZS execution time, and an average reduction of 63.58% in the AME execution time, resulting in a small average BD-BR efficiency loss of only 0.90%. These results are competitive compared to related works in the literature and demonstrate that the strategy of using machine learning to reduce the computational cost of VVC has the potential to continue yielding significant results in future solutions.