Performance analysis of machine learning for arbitrary downsizing of pre-encoded HEVC video

摘要

Nowadays, broadcasters deliver ultra-high resolution video to their consumers. This live video is sent to a set-top box for display on a television. However, if one or more users in the home want to view the same video on their personal mobile devices with a lower display resolution and limited processing power, decoding the original ultra-high resolution video would result in stuttering and quickly drain the battery life on these devices. To enable a satisfactory consumer experience, the resolution of the video stream should be adapted to the target mobile device at the set-top box. The aim of this paper is to investigate the performance of different machine learning strategies to arbitrary downsize video pre-encoded with the high efficiency video coding standard (HEVC). These machine learning techniques exploit correlation between input and output coding information to predict the splitting behavior of HEVC coding units. Several machine learning algorithms are optimized. Additionally, both online and offline training strategies are tested. Of the tested algorithms, online-trained random forests achieve the best compression-efficiency with a bit rate increase of 5.4% and an average complexity reduction of 70%1.