Machine-Learning Based High Efficiency Rate Control for AV1

Yi Chen, Yunhao Mao, Shiqi Wang, Xianguo Zhang, Sam Kwong

Abstract: Recent years have witnessed the increasing demand of video coding technologies, which have been continuously developed to meet various requirements in video-related applications. Developed by Alliance for Open Media (AOM), the AOMedia Video 1 (AVl) is an open-source and royalty-free standard. Herein, we achieve high efficiency rate control for AVI based on the machine-learning model, which establishes the rate-quantization relationship in a data-driven manner. More specifically, the Supporting Vector Regression (SVR) is used for rate model parameter estimation. The model is trained using sufficient training data, and incorporated in the encoder. Compared to the default rate control scheme in AV 1, experimental results have shown that 2.01% bitrate could be saved with tolerable bitrate error.