Desenvolvimento de Soluções Baseadas em Aprendi zado de Máquina para Transcodificação Rápida de Vídeo ao Formato AOMedia Video 1.

Abstract

Video encoders are currently very important tools for enabling common applica tions in our daily lives, whether in applications dedicated to video transmission for entertainment, such as YouTube and Netflix, or in social networks, such as Instagram or TikTok, or even for private communication, such as on video calls. Even with the use of efficient video encoders, video content represents a considerable portion of the world’s data traffic over the Internet. For this reason, this is an area of unique relevance in the scientific community and the definition of new standards and video compression formats has been a constant. Considering the large number of video formats and standards, modifying encoded video files for different purposes is a common practice, whether to provide compatibility between devices or even to adapt a coded video to different situations, such as bit rate adaptation and resolution change. This modification is called video transcoding and has several applications. One of the applications, called heterogeneous transcoding, tipically involves updating videos encoded in an older format to a more recent one with greater compression efficiency. However, this task requires a significant computational effort, as it involves decoding and re-encoding the video in sequence. Therefore, part of the scientific community active in the video coding field has been looking for solutions to speed up the video transcoding process. This thesis is focused on this goal. The thesis initially presents the state of the art in video transcoding, its applications and techniques. In this thesis, seven fast transcoders are proposed for the AOMedia Video 1 (AV1) format, based on other formats widely used by the video streaming industry. Among the proposals, those that employ predictive models trained by machine learning algorithms to accelerate the encoder partitioning decisions stand out. In order to enable the agile development of fast transcoders, this thesis also proposes a processing pipeline, which allows, among other things, the automation of training predictive models and the scheduling of experiments to test them. As a proof of concept for this methodological proposal, five fast video transcoders were developed with the pipeline, all of them employing decision tree models. The obtained results indicate that it is possible to accelerate the transcoding process for the AV1 format between 12% and 55%, with losses in coding efficiency that vary between 0.96% and 12.85%, depending on the source format.