AVS To HEVC High Efficiency Video Transcoding Technique

With the wide application and fast development of network and multimedia technology, it becomes a trend to transmit videos through the network. However, the bandwidth is limited, and the video data needs to be compressed. To meet this demand, many video coding standards came out, including MPEG-4, MPEG-2, H.264, AVS and HEVC. Video resources are various, and terminal equipment have different abilities to process video streams, such as display ability and storage ability. Considering this, video transcoding techniques can be used to solve the problems.HEVC is the latest video coding standard, which can improve about 50 percent of compressing efficiency compared with H.264. AVS was developed by China and palyed a siginificant role in Chinese markets. To make AVS more compatible with international standards, we studied the AVS to HEVC high effiency video transcoding techniques. And the topic contains two basic ideas, which are transcoder based on ROI detection using visual characteristics and transcoder based on machine learning.One of the methods is AVS to HEVC video transcoder based on ROI detection using visual characteristics. This method takes account the human visual system into the transcoder. Combined with this, the transcoder can achieve better subjective video quality. The algorithm is mainly composed of two parts: ROI detection algorithm and fast transcoding algorithm. Firstly, we extract the coding information contained in AVS video stream. The information includes prediction mode, motion vector and transform coefficients. With the coding information, we part the video into three kinds of regions: region of most-interest(most ROI), region of interest(ROI), region of less-interest(less ROI). After then, the detection results can be used to guide the re-encoding process in HEVC. For different regions, algorithms with different complexity are used accordingly. In this way, the computational complexities can be decreased dramatically. Experimental results show that the algorithm can save about 50% of transcoding time, with PSNR drop less than 0.05 d B, while maintaining better visual quality.The other method is AVS to HEVC transcoding based on SVM. We apply the machine learning into transcoder. This algorithm divided the transcoding process into two stages: training stage and transcoding stage. Firstly, we also extract the coding information including prediction mode, motion vector and transform coefficients, from the input video stream. The information can form feature vectors for training. Using machine learning, we obtained the mapping relationship between feature vectors and CU split flag in HEVC. In transcoding stage, the training model is used to predict the CU partition. Combined with simple fast mode selection methods, the proposed algorithm can achieve good transcoding efficiency. Experimental results showed that with this algorithm, the transcoder can speed up about 60%~70%.