Study On Spatial Downscaling Video Transcoding

With the development of modern multimedia communication technology, there is a strong need for users to access videos originally captured in a high resolution through the mobile multimedia capable devices. Thus, a transcoding mechanism is required to make the content adaptive to the capabilities of diverse client devices (e.g. PDA, smart phone, etc).Due to the limit of network bandwidth and display material of terminal device, the high definition source video encoding bit-rates should be transformed into lower resolution definition formats which adapt to the receiving devices.H.264, as the emerging video coding standard, achieves much higher coding efficiency than previous standards. So it is being widely used in modern video market. It's very popular to research H.264 downscaling transcoding technique recently.This thesis discusses the key techniques of H.264 downscaling transcoding, including macroblock re-encode mode selection, motion vector re-estimation and so on. Re-encoding mode selection algorithm includes intra and inter cases. Because full search motion estimate algorithm takes too much time during predicting the inter image, it is essential to reuse the motion information from the input bit-rates. Further more, the motion refinement method include the increase of re-encoding bit-rates while the prediction is improved. This thesis reviews the traditional method of downscaling video transcoding and proposes a new algorithm on H.264 downscaling transcoding, which uses improved measure to predict motion vectors, and uses an adaptive method to judge the re-encoding mode of a video sequence.In case of exquisite sequences, the transcode picture quality of the proposed method is about 1dB higher than those of traditional MV composition methods, and 0.1～0.7dB higher than that of DIM method, while makes the same bit-rates. In case of mild sequences, it produces only 30%～60% bits than DIM method, while preserving high picture quality.In the last part of the thesis, the author introduces another work which was done during the master state: implement of AVS entropy decoding and optimization on DSP. Using instruction level and algorithm level optimization, this method improves the parallelism degree of DSP. The entropy decode speed improved by more than 3 times, which contributes to the achievement of real-time decoding.