MPhil Thesis Defence


"Fast Video Transcoding and Graphics Hardware Acceleration"

By

Mr. Chi Wang Ho


Abstract

Video transcoding is one of the key technologies in modern multimedia
systems to enable effective access for a wide diversity of devices and
formats. In applications that deliver live videos over bandwidth limited
channel to multiple users, a transcoder is needed to provide dynamic
adaptations in the bit-rate and the format of the video bit-stream in
order to serve different users simultaneously. However, a robust and
flexible transcoder usually requires a cascaded decoding and re-encoding,
which is computational expensive. Thus, it is important to re-use the
information of the input video bit-stream as efficient as possible to
reduce the complexity while maintaining high output quality.

In addition, the transcoding process can be further speed up by making use
of computer graphics hardware. Graphics processing units (GPUs) provide
tremendous memory bandwidth and computational horsepower, and
general-purpose computation on GPUs (GPGPU) has demonstrated a great
success in recent years. Some researchers have used GPU for image and
video processing, such as transformation, motion compensation and motion
estimation, etc. Further research on using GPU for video processing, in
particular video encoding, is needed to exploit the possibilities of speed
improvement.

This thesis is divided into two parts. We first address the
algorithm-level solutions for fast transcoding by re-using the information
extracted from the input video bit-stream. Two rate control schemes are
proposed. The first proposed method, namely Zero-Residue Pre-Selection
(ZRPS), reduces the computational complexity of a well-known rate control
scheme TMN-8 by selecting a sub-set of macroblocks for the calculations.
The rate information of the input video bit-stream is used to select these
macroblocks. Another rate control scheme, Low-Complexity Rate Control
(LCRC), is proposed to use the relationship between the rate of the input
video bit-stream and the output video bit-stream on a per-row basis. Since
no explicit complexity measurement of macroblocks is needed, it is good
for H.264/AVC under the rate-distortion optimization framework. In
addition, we also propose a motion vector refinement scheme - Minimum Cost
Tendency Search (MCTS) for transcoding from H.263 to H.264/AVC. MCTS
explicitly considers the difference between their cost functions for
motion estimation and results in a more stable performance over a wide
range of output quality. In the second part of this thesis, a GPU-based
motion estimation for H.264/AVC is presented. It is flexible enough that
can apply on different graphics chips and obtain a good performance with
simple fine tuning.


Date:				Tuesday, 15 August 2006

Time:				10:00a.m.-12:00noon

Venue:				Room 5501
				Lifts 25-26

Committee Members:		Dr. Oscar Au (ECE, Supervisor)
				Dr. Gary Chan (Supervisor)
				Dr. Qian Zhang (Chairperson)
				Dr. Philip Fu


**** ALL are Welcome ****