Improved standard-conforming video transcoding techniques.

摘要

Many video coding standards have been established for various video applications, such as H.263 for low-bit-rate two-way video communications, MPEG-2 for broadcasting and general high-quality video applications, and MPEG-4 for streaming video and interactive multimedia applications. As digital video applications become more popular, increasing amounts of video content are encoded in different formats. Consequently there are increasing needs to dynamically convert the video between different formats. A video transcoder is a device that converts video from a compressed format into another compressed format. It is one of the enabling technologies for video delivery over heterogeneous networks. The goal of transcoding research is to achieve the best possible quality-complexity tradeoff given application constraints. The key to achieve this goal is to intelligently utilize the coding statistics of the input compressed video stream to facilitate the transcoding process.; In this dissertation, we propose several techniques to improve transcoding performance. First, we present an efficient motion re-estimation scheme that can handle not only the temporal/spatial resolution conversion, but also the interlaced video processing. Exploiting the spatial and temporal correlations between the input motion vectors, we derive the motion vectors for the transcoded output video from appropriate input motion vectors. Compared with the straightforward, cascaded decoder-encoder implementation using full-search motion estimation, the proposed algorithm achieves virtually the same quality with greatly reduced computational requirement. We then present an adaptive picture-layer bit-allocation algorithm based on estimated output picture complexities. It achieves better bit-allocations to different types of pictures, and handles scene-changes better than conventional approaches. It the end, we propose an operational rate-distortion optimized macroblock-layer rate-control algorithm based on the Lagrange Multiplier (LM) algorithm. The LM algorithm is a quasi-optimal solution to this rate-control problem, but is computationally too expensive for practical applications. Taking advantage of the properties of the low-bit-rate video and the LM algorithm, we propose an approach to reduce its computational complexity. The proposed algorithm outperforms H.263 TMN8 in terms of video quality and buffer behavior with slightly increased computational requirement.