Deep Motion Vector Prediction for Versatile Video Coding

Abstract

Versatile video coding (VVC) has made significant progress in the compression efficiency of video coding. It achieves better performance and halves bitrate compared to the high efficiency video coding (HEVC) under the same visual quality. However, the complexity of VVC has also been significantly increased, especially in inter prediction. This paper proposes an improved motion vector prediction method based on neural networks for motion estimation (ME). Firstly, dynamic weights are proposed in the process of selecting the best MVP for advanced motion vector prediction (AMVP); secondly, we build the motion vector prediction model based on the deep neural network; finally, the model is embedded in VVC to acquire a more accurate MVP and reduce the encoding complexity of ME. Experimental results show that the proposed algorithm can reduce the encoding time of motion estimation under the premise of guaranteeing video quality.