Non-iterative Content-Adaptive Distributed Encoding Through ML Techniques

Distributed encoding is desired in content preparation workflows in the cloud to reduce turnaround times. Content-adaptive bit allocation at title, chunk, and even frame level, as opposed to using a fixed ladder of content-independent bit-rates and resolutions, have been proposed to achieve efficiencies in storage and delivery. Many of these methods tend to be iterative in nature and consume significant additional compute resources over 2-pass Variable bit-rate (VBR) encoders. With live use-cases attempting to use a set of distributed encoders to achieve higher bit-rate savings (both through use of higher compression presets and through content-adaptive bit allocation) and consistent quality, there is a need to limit the increase in computational complexity. In this paper, we propose a non-iterative codec-agnostic approach that employs machine learning techniques to perform consistent quality content-adaptive encoding within the constraints of a maximum bit-rate in a manner that makes it equally suitable for live and on-demand workflows. The method has the ability to take a target subjective rating level and allocate appropriate bits for each group of frames to achieve that. The proposed approach also anticipates automatic selection of the right resolution and frame-rate for a given representation within an ABR set, and content-specific encoding parameters to maximize the bit savings and/or visual quality. Test results are presented over a wide range of content types comparing the performance of the proposed approach against 2-pass VBR methods. Initial results indicate that the proposed approach can recover ~85% of the bit-savings possible with exhaustive techniques. The computational complexity of the proposed approach is only 15-20% of 2-pass VBR encoding.

Published

2017-10

Content type

Original Research

Keywords

Distributed encoding, content-adaptive bit-rate/resolution ladder for adaptive bit-rate (ABR) streaming, HLS, DASH, machine learning (ML), consistent quality encoding, non-iterative, live and on-demand over-the-top (OTT) streaming, content storage/delivery efficiency

DOI

10.5594/M001783

ISBN

978-1-61482-959-1