CTU Partition for Intra-Mode HEVC using Convolutional Neural Network
Hari Pattimi, Vivek Jadhav, B.K.N Srinivas Rao · 2022 IEEE International Symposium on Smart Electronic Systems (iSES) · 2022
High Efficiency Video Coding (HEVC / H.265) decreased bit rates by 50% compared to the Advanced Video Coding (AVC / H.264) standard. The HEVC Coding Tree Unit (CTU) partition uses a quadtree structure. This structure improves coding unit performance by letting each Coding Unit (CU) be divided into four smaller CUs. The use of brute force search for Rate Distortion Optimization (RDO) in CTU partition makes HEVC encoding very complicated. This paper proposes a Convolutional Neural Network (CNN) that predicts CTU partition instead of RDO for intra-mode HEVC, which makes HEVC encoding simpler. CNN gets 64×64 CTU as input, and the prediction of depth for each 64×64 coding unit is represented by a 16×16 matrix of each 4×4 block. If redundant information in a 16×16 matrix can be reduced to a 1×16 vector, then the output of a 64×64 matrix can also be represented as a 1×16 vector. CNN model is integrated with HM reference software and compared to original HM software in terms of encoding time, BD-BR, and BD-PSNR to find out how much the encoding process has become simpler. The proposed HMCNN framework reduces the computational complexity of intra mode HEVC by up to 70.39 percent at the cost of a 4.27% drop in bitrate.