Transcodage efficace de MPEG-4 partie 2 à H.264 basé sur les modes de codage, les vecteurs de mouvement et l'information résiduelle de la source vidéo MPEG-4 partie 2

Abstract

The MPEG-4 part 2 standard developed in 1995 by ISO MPEG is widely used in today’s multimedia applications, such as broadcasting, streaming, mobile video and video games. However, H.264 (also called MPEG-4 part 2 or MPEG-2 AVC), the latest video standard developed jointly by the ISO and the ITU in 2003, offers better compression performances than MPEG-4 part 2. Its visual quality is comparable to that of MPEG-4 part 2, but at a compression rate two to three times higher. While an increasing number of applications is adopting this new standard, the co-existence of the two standards leads to interoperability problems. MPEG-4 part 2 to H.264 transcoding has become a necessity not only to enable communication between terminals supporting these two standards, but also for platforms that want to convert MPEG-4 part 2 content to H.264 to benefit from its improved compression factor. Unfortunately, H.264 achieves its high-efficiency compression rate at the expense of high computational complexity, making MPEG-4 part 2 to H.264 conversion unsuitable for realtime applications. We have developed two efficient transcoding algorithms, with the aim of reducing such computational complexity while maintaining good visual quality. We extracted the coding modes (CMs), motion vectors (MVs) and residual information during the MPEG-4 part 2 decoding phase, and used them in the H.264 compression phase to reduce the latter’s computational complexity. In particular, we have exploited several properties of residual information, for example, the relationship between macroblock (MB) partitioning and residual information. Also, we use the frequency distribution of H.264 CMs as a function of MPEG-4 part 2 CMs to classify the MBs and eliminate less probable H.264 candidate CMs. This significantly reduces the set of candidate CMs to test, while preserving good visual quality. Our algorithm innovates again, by taking advantage of the properties of the residual information in the motion estimation phase, where the efficiency of MVs is measured. This allows us to refine only the inefficient ones, thereby avoiding computationally expensive MV refinements that would not contribute to improving visual quality. Furthermore, we propose to use a relative measure of the residual information and exploit the correlation between successive frames to make our transcoding methods adaptive to bit rates and video characteristics. Doing so has allowed us to obtain efficient transcoding for a wide range of bit rates and resolutions. Our methods have been tested and compared with state-of-the-art transcoding methods using over fifty video tests, which cover various resolutions from QCIF (176×144 pixels) to HD (1920×1080), running each for five or six bit rates. We obtain speed-ups 3× to 5× those of the cascade approach (decode/re-encode), with quality losses of 0.15 dB to 0.5 dB, on average, for QCIF using Intel IPP codecs, which are highly optimized for speed. With the exception of one method that yields a higher speed-up, but at a huge quality loss (up to 4 dB), the proposed algorithms lead to significantly better results, in terms of both speed-up and quality, than the state-of-the-art methods.

Date	24 Jan 2012
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Stéphane Coulombe (Supervisor)