BIDIMENSIONAL DICTIONARY AND CODING SCHEME FOR A VERY LOW BITRATE MATCHMG PURSUIT VIDEO CODER
I Io · 2004
In this papa we present a video coder hascd on a redundant dictionary that employs non szparahle bidimensional functions. 'The algorithm used for the reprcscntation of the residual Ol' motion estimation is Matching Pursuit (MP). 'Thc proposed dictionai) has new features capahlc of catching curved and oriented contours typical or thc residual ut' motion estimation in natural video scquences. An adaptive grid approach is adopted for the coding of atoms and this is combincd with an innovative coding method employing an atithmetic encoder. The proposed coder shows improvements over 13.264 in terms of compression efficiency, with up to 20% gains for vety low bitrates. 1. INTRODUCHON Mobile vidco peer-to-per communications and mobile video streaming are dramatically changing the way we think about communications. In this environment, especially in 3G networks for which mor resiliency is directly provided at the data link layer, the saving of bandwidth is the key issue. Since the late 80ies and the appearance of H.261 [I], video comprcssion efficiency has kept on evolving at a steady rate. Remarkable improvements have been achieved for each component of the coding engine, though the basic architecture, based on block matchingdct-vlc remained almost the same, including the very recent baseline profile version of H.264 121. In this paper we propose a video coder based on MP. Diffmntly from the previous MP dictionaries, the one introduced here adopts 2d functions (also called atoms) able to catch oriented edges in the displaced frame difference (dfd) images, namely those images that are the residual signal of the motion estimation. Atoms are arranged on a variable grid and coded using an arithmetic coder whose contest adapts to the characteridics of the structure of the source. Results show an improvement of up to 20% when compared with H.264 for bitrates in the range of 20-40 kbs.