Redução de consumo energético para transformadas do padrão versatile video coding com auxílio de aprendizado de máquina supervisionado

Abstract

The Versatile Video Coding standard was released by the Joint Video Experts Team in 2020, introducing several tools aimed at improving compression efficiency compared to previous standards. One of the innovations is the Multiple Transform Selection, which enables the encoder to choose between different types of transforms to bet ter adapt to the local characteristics of the video signal. Multiple Transform Selection employs the discrete cosine transform type II , discrete cosine transform type VIII, and discrete sine transform type VII, also allowing distinct combinations of these transforms in the horizontal and vertical directions. Although this flexibility improves compression efficiency, it also significantly increases computational complexity, since multiple com binations of transforms and block sizes must be evaluated by the encoder. In light of these challenges, this thesis proposes the development of dedicated hardware ar chitectures for the Multiple Transform Selection module of the Versatile Video Coding encoder, focusing on reducing energy consumption and enabling real-time compres sion. The project is structured in three main stages: a detailed analysis of Multiple Transform Selection usability in the Versatile Video Coding reference software, the in tegration of machine learning-based predictive models into the encoding flow, and the implementation of optimized hardware architectures based on data extracted from the encoder. To reduce the complexity of the transform selection process, predictive mod els were developed using machine learning algorithms. These models were trained on data directly extracted from the Versatile Video Coding reference encoder, and their purpose is to anticipate which transforms are most likely to be selected in each sit uation. This prediction enables unnecessary transforms to be disabled, reducing the number of combinations tested and consequently lowering processing time and en ergy consumption. The proposed hardware architecture was designed to support both the traditional software execution flow and the modified flow that incorporates predic tive models. The thesis presents the methodology for feature extraction and selection, model training, integration into the VTM encoder, and energy and area results across different configurations and video resolutions. The results demonstrate that, despite a slight increase in bitrate, the adoption of predictive models led to significant reductions in encoding time and energy consumption. The tests demonstrate that, even with a small loss of 0.89% in coding efficiency, the adoption of predictive models resulted in significant reductions, averaging up to 7.98%, in processing time when implemented in the reference software. Additionally, the potential implementation of predictive mod els in hardware is discussed, using simple conditional structures, which allows their efficient integration into embedded systems with limited computational resources. The proposed hybrid approach, combining machine learning and optimized hardware archi tecture, represents a promising strategy for enabling energy-efficient Versatile Video Coding encoders, achieving energy consumption reductions of up to 71.37% for 4K resolutions. This contribution is relevant for both portable device applications and real time, high-resolution video compression scenarios