Fast QTMT for H.266/VVC Intra Prediction using Early-Terminated Hierarchical CNN model

Versatile Video Coding (VVC) has been standardization in July 2020. Compared to previous High Efficiency Video Coding (HEVC) standard, VVC saves up to 50% bitrate for equal perceptual video quality. To reach this efficiency, Joint Video Experts Team (JVET) has introduced a number of improvement techniques to VVC model. As a result, the complexity of VVC encoding also greatly increases. One of the new techniques affects to the growing of complexity is the quad-tree nested multi-type tree (QTMT) including binary split and ternary splits, which lead to a block in VVC with various shapes in both square and rectangle. Based on the aforementioned information we propose in this paper a new deep learning based fast QTMT method. We use a learned convolutional neural network (CNN) model namely Early-Terminated Hierarchical CNN to predict the coding unit map and then fed into the VVC encoder to early terminate the block partitioning process. Experimental results show that the proposed method can save 30.29% encoding time with a negligible BD-Rate increase.