Optimal DSP Based Integer Motion Estimation Implementation for H.264/AVC Baseline Encoder

Optimal DSP Based Integer Motion Estimation Implementation for H.264/AVC Baseline Encoder

Imen Werda1, Haithem Chaouch1, Amine Samet2, Mohamed Ben Ayed2, and Nouri Masmoudi1
1 National School of Engineering, University of Sfax, Tunisia
2 High Institute of Electronics and Communication, University of Sfax, Tunisia

Abstract: The coding gain of the H.264/AVC video encoder mainly comes from the new incorporated prediction tools. However, their enormous computation and ultrahigh memory bandwidth are the penalties. In this paper we present an approach supporting efficient data reuse process to avoid unnecessary memory accesses and redundant motion estimation computations combined with a novel fast algorithm. A merging procedure joining search origin, search pattern and new variable block size motion estimation for H.264/AVC is detailed in this paper. Those approaches yield good tradeoffs between motion estimation distortion and number of computations since they invest and exploit the centre-biased characteristics of the real world video sequences: a reliable predictor determines the search origin, localizing the search process. An efficient search pattern exploits structural constraints within the motion field. A new fast block size selection DSP-based algorithm allows simultaneous fidelity of the video quality and the reduction of the computational cost. Extensive experimental work has been done, results of which show that our approach gives a speed up of 1.14 times over that of the recent fast algorithms and 10 times over the spiral search algorithm on average, with a negligible degradation in peak signal-to-noise ratio. In addition, interesting memory bandwidth is further saved with the proposed data reuse techniques at architecture level.

Keywords: H.264/AVC, search centre, block matching algorithm, pattern search, variable block size, complexity, video quality, PSNR, SSIM.

 Received March 7, 2008; accepted September 1, 2008

 

Full Text

 
Read 4784 times Last modified on Wednesday, 20 January 2010 00:24
Share
Top
We use cookies to improve our website. By continuing to use this website, you are giving consent to cookies being used. More details…