Coding unit partition prediction technique for fast video encoding in HEVC

摘要

In this paper, a prediction technique for the best coding unit (CU) partition tree is proposed to avoid the repeated computation on the variable-sized prediction units (PU) and transform units (TU) in high efficiency video coding (HEVC). In HEVC, many advanced coding tools are introduced to decrease the bit-rate of coded videos with higher visual quality. Of these tools, many seek to partition each CU in the incoming frame into a quad-tree structure of variable sized blocks while performing prediction and transform computations to obtain the optimal rate-distortion performance. During partition tree generation, the CU block and its four partitioned child-blocks need to become PU and TU for calculating the coding rate-distortion cost. This process requires more execution time from the processor because of highly repeated computations on PUs and TUs before finding the best quad-tree partition tree. The proposed prediction technique is based on edge information from the content of the considered CU and the depth information of partition tree from the co-located CU in the neighboring video frames for being able to adapt the reduction of computation time with the varying content in the incoming frames. When the CU is in a smooth frame, the information from the neighboring frames would dominate the prediction results of the partition tree; however, the edge information of the current CU mainly determines the partition tree. With the prediction technique, experiments on several benchmark videos show that the encoding time of HEVC can be reduced by 44.1 % at an acceptable degradation of coding efficiency.