A Machine Learning Approach for Fast CU Partitioning and Time Complexity Reduction in Video Coding

Abstract

Modern video codecs have high capability to minimize the bits and also have an immense complexity because of testing more combinations during the RDO (Rate Distortion Optimization) process. Partitioning a coding unit (CU) into smaller ones is the leading cause of the increased time complexity. For this reason, a fast and efficient algorithm is needed for real-time applications. In this paper, an approach for reducing the time complexity of CU splitting using a decision tree classifier is proposed. If a CU contains only a single motion, i.e. if the CU falls in a homogeneous region, this CU can be skipped from further partitioning at the time of RDO process. Experimental results show that on average 45.02% time complexity reduction is possible by implementing the skip criterion. However, as a trade-off on average 1.38% bit rate is increased over standalone HEVC.