Fast Inter Partitioning of CU Based on Neural Network for Versatile Video Coding
Xudong Zhang, Jing Chen, Debo Zhang, Jiaxin Zeng, Yuting Zuo, Quanxu Zhao, Wanjian Feng · 2024
The versatile video coding standard (H.266/VVC) is able to save about 50% bit rate on average compared to the previous generation video coding standard H.265, but at the cost of a huge increase in coding complexity. In order to reduce the time consumption of VVC during inter-frame coding, A method based on neural network which does not contain convolutional operations is proposed. The network is trained for 20 sizes and 5 levels of CU using residual image and some temporary variables and the network is invoked during the video encoding process and skip several partitioning modes with lower probability to speed up. The network structure contains only fully connected layers and Leaky ReLU layers, which can fully utilize the CPU performance. Compared to the original scheme, our method saves 60.07% coding time at the cost of 4.41% bitrate in Random Access configuration.