Fast CU partition-based machine learning approach for reducing HEVC complexity

With the development of video coding technology, the high efficiency video coding (HEVC) provides better coding efficiency compared to its predecessors H.264/AVC. HEVC improves rate distortion (RD) performance significantly with increased encoding complexity. Due to the adoption of a large variety of coding unit (CU) sizes, at RD optimization level, the quadtree partition of the CU consumes a large proportion of the encoding complexity. Hence, the computational complexity cost remains a critical issue that must be properly considered in the optimization task. In this paper, two machine learning-based fast CU partition method for inter-mode HEVC are proposed, to optimize the complexity allocation at CU level. First, we propose an online support vector machine (SVM)-based fast CU algorithm for reducing HEVC complexity. The later was trained in an online way. Second, a deep convolutional neural network (CNN) is designed to predict the CU partition, in which large-scale training database including substantial CU partition data is considered. Experimental results demonstrate that the proposed online SVM can achieve a time saving of 52.28% with a degradation of 1.928% in the bitrate (BR). However, the proposed deep CNN can reduce the encoding time by 53.99% with 0.195% BR degradation. Compared to the state-of-the art, the two proposed approaches outperform the related works in terms of both RD performance and complexity reduction at inter-mode.