Please use this identifier to cite or link to this item: http://hdl.handle.net/10397/83170
Title: Efficient techniques for video rate control
Authors: Lo, Kam-fai
Degree: M.Phil.
Issue Date: 2005
Abstract: Since the first introduction of digital video compression, the development of the compression technology has never stopped. The increasing demand for digital video applications puts increasing pressure on various intrnational standards such as MPEG-1, MPEG-2, MPEG-4, H.263 and H.264 to keep improving. Rate control is one of the most important responsibilities of a video encoder, especially for video transmission. The output rate of a video must be controlled according to the channel bandwidth and the buffer constraints; this can be achieved by adjusting the quantization parameters used to compress the video frames. These quantization parameters are used to control the trade-off between the video quality and the output bit rate. Although rate control is important, it is not defined in the intrnational video standards since it does not affect the design of the decoders. Therefore, video applications must have their own control scheme. The reference models of the existing standards employ different rate-control schemes. In H.263+, the scheme is based on the classical logarithmic rate-distortion function. The rate model of the recent standards - MPEG-4 and H.264 - employs quadratic rate-distortion functions. These schemes rely on a statistical rate model derived from the information theory, while some other schemes are based on fuzzy logic. Like the classical rate model, the fuzzy-logic rate controller needs the statistical data of the video in order to estimate suitable quantization parameters. However, the fuzzy-logic approach usually requires simpler implementation as compared to those based on the rate models. All these rate-control schemes are designed for real-time video coding, and their computations should be relatively low. Some rate-control schemes also consider rate-distortion optimizations, which need an iterative process to search for the optimal or suboptimal combination of the quantization parameters for predictive coded frames. Such schemes always outperform the real-time control schemes in terms of rate and quality, but a larger amount of computations are required. Hence, these schemes are only suitable for non-real-time applications. In this research, we explore the most popular rate-control schemes, including those adopted in the reference software of H.263+ (TMN8), MPEG-4(VM5.0) and H.264(JM8.1a), the fuzzy-logic rate controller, and the rate-distortion optimization schemes. We also propose a new rate-control scheme based on a distortion-rate model, which can be extended easily to a higher order for different coding algorithms without requiring much effort to determine the quantization scales. The formulation of the cubic model is derived from the quadratic rate-distortion model and from simple curve fitting. Our model is implemented in H.264 JM8.1a and its performance is compared to JM8.1a with different test sequences and bit rates. Experimental results show that our proposed model outperforms the quadratic model in terms of both PSNR and bit rate. The second part of the research aims at reducing the computations of an existing rate-distortion optimization scheme. The computational effort is reduced to half with only a slight decrease in video quality. The cubic model scheme is implemented in the Joint Model 8.1a (JM8.1a) of H.264, so that the performance can be evaluated. Experimental results show that our scheme can achieve better performance leels than JM8.1a. A study of the typical, basic hybrid coders and the new features of H.264is also conducted in this research.
Subjects: Hong Kong Polytechnic University -- Dissertations.
Digital video.
Video compression.
MPEG (Video coding standard)
Multimedia systems.
Pages: xi, 112 leaves : ill. ; 30 cm.
Appears in Collections:Thesis

Show full item record

Page views

2
Citations as of May 22, 2022

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.