Название: Versatile Video Coding (VVC): Machine Learning and Heuristics
Автор: Mario Saldanha, Gustavo Sanchez, Cesar Marcon, Luciano Agostini
Издательство: Springer
Год: 2022
Страниц: 128
Язык: английский
Формат: pdf (true), epub
Размер: 23.2 MB

This book discusses the Versatile Video Coding (VVC), the ISO and ITU state-of-the-art video coding standard. VVC reaches a compression efficiency significantly higher than its predecessor standard (HEVC) and it has a high versatility for efficient use in a broad range of applications and different types of video content, including Ultra-High Definition (UHD), High-Dynamic Range (HDR), screen content, 360º videos, and resolution adaptivity. The authors introduce the novel VVC tools for block partitioning, intra-frame and inter-frames predictions, transforms, quantization, entropy coding, and in-loop filtering. The authors also present some solutions exploring VVC encoding behavior at different levels to accelerate the intra-frame prediction, applying statistical-based heuristics and Machine Learning (ML) techniques.

Chapters 6–11 present heuristic and Machine Learning solutions we developed for providing encoding time reduction in VVC intra-frame prediction with minimal impact on coding efficiency. Chapter 6 presents a fast decision scheme based on a heuristic for the block partitioning of luminance blocks. This scheme explores the selected intra-frame prediction modes and the variance of block samples to decide when to avoid some block size evaluations. This solution reduced the encoding time by 43.23% at the cost of a 0.80% Bjontegaard Delta Bit Rate (BD-BR) [26] increase, which is a metric used to evaluate the coding efficiency losses. A configurable fast block partitioning solution based on machine learning for luminance blocks is presented in Chap. 7. In this case, a machine learning classifier called Light Gradient Boosting Machine (LGBM) was used to reduce the number of evaluated partitions. A set of LGBM classifiers was trained and validated using real digital video sequences, this solution reduced the encoding time from 35.22 to 61.34% with BD-BR increase from 0.46 to 2.43%.

Chapter 8 presents our Machine Learning approach to generate a fast decision scheme for intra-mode selection of luminance blocks. This scheme was modeled with two solutions using decision trees and one solution based on a heuristic. This scheme provided an encoding time reduction of 18.32% with a 0.60% BD-BR increase. A machine learning solution was also developed to reduce the encoding time for intra-transform mode selection of luminance blocks and this solution is presented in Chap. 9. This solution was also modeled using two decision tree classifiers, reaching an encoding time reduction of 11% with an increase of 0.43% in BD-BR. The last solution presented in this book is a heuristic-based fast block partitioning scheme for chrominance blocks.