A CNN-based Coding Unit Partition in HEVC for Video Processing

Abstract

The High Efficiency Video Coding (HEVC) standard is able to provide efficient video encoding for broadcast mobile applications. HEVC employs the Coding Tree Unit (CTU) to improve intra coding performance. The partition of each CTU is determined by calculating the Rate Distortion (RD) cost in an exhaustive search process manner, and the calculation results in high computation complexity. In this paper, we propose a deep learning approach for Coding Unit (CU) partition in HEVC intra coding. The proposed approach utilizes the Convolutional Neural Network (CNN) to predict CU partition during the CU search process. The input of CNN is a CU while the output is a partition flag, which refers to partition the current CU or not. The proposed approach employs Graphics Processing Unit (GPU) to accelerate the CNN computation. The experiment results demonstrate the proposed approach achieves 63.19% and 66.01% encoding time reduction on average with negligible perceptual quality loss.