CU Size Decision for Low Complexity HEVC Intra Coding based on Deep Reinforcement Learning

High efficiency video coding (HEVC) uses a quadtree-based structure for coding unit (CU) splitting to effectively encode various video sequences with different visual characteristics. However, this new structure results in a dramatically increased complexity that makes real-time HEVC encoding very challenging. In this paper, we propose a novel CU size decision method based on deep reinforcement learning and active feature acquisition to reduce HEVC intra coding computational complexity and encoding time. The proposed method carries out early splitting and early splitting termination by considering the encoder and CU as an agent-environment system. More specifically, through early splitting, the proposed method precludes the need for rate-distortion optimization at the current level. In addition, through early splitting termination, it disposes of the lower level computations. The proposed method provides a very fast encoder with a small quality penalty. Experimental results show that it achieves a 51.3% encoding time reduction on average with a small quality loss of 0.041 dB for the BD-PSNR, when we compare our method to the HEVC test model.