Transcodage rapide de H.264 à HEVC basé sur la propagation du mouvement et une traversée postfixe des unités de codage arborescent

Abstract

In 2013, the Joint Collaborative Team on Video Coding completed HEVC, the most recent video compression standard. Compared to its predecessor, H.264, the new standard HEVC can save approximately 50% of the bitrate for similar video quality. To take advantage of HEVC coding efficiency and to ensure systems interoperability, several H.264 sequences must be transcoded to HEVC. The simplest video transcoding approach, called cascade pixel-domain transcoding (CPDT), decodes the input sequence entirely and re-encodes the pixel data in the output format. This approach achieves high coding efficiency and offers great flexibility on video encoding parameters. However, it is very complex computationally. To reduce this complexity, several approaches reuse extracted information – such as coding modes, motion information and encoded residuals – from the incoming bitstream to speed up the HEVC encoding process. For inter frames, these approaches typically focus on fast mode decision and fast motion estimation. Most of these approaches preserve the coding efficiency, but achieve limited speedup (usually between 2 and 4x). In this thesis, we propose a faster H.264 to HEVC transcoder that also preserves the coding efficiency. Our solution is composed of a motion propagation algorithm and a fast mode decision framework. The motion propagation algorithm creates a motion vector candidate list at the coding tree unit (CTU) level and, thereafter, selects the best candidate at the prediction unit level. This method eliminates computational redundancy by pre-computing the prediction error of each candidate at the CTU level and reusing the information for various partition sizes. The fast mode decision framework is based on a post-order traversal of the CTU, and includes several mode reduction techniques. In particular, the framework permits the early termination of the rate-distortion cost computation, a highly complex task, when a mode is unpromising. Moreover, a method is presented to recursively determine, with the help of H.264 modes and information created by the motion propagation algorithm, whether or not each coding unit must be split. This allows the pruning of unpromising sub-partitions. Compared to a CPDT approach, the experimental results show that the proposed solution is on average 7.81 times faster, for an average BD-Rate of 2.05%. Our experiment shows that these results exceed those of state-of-the-art methods.

Date	19 Dec 2016
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Stéphane Coulombe (Supervisor)