A fast CU partition algorithm of ERP 360-degree video based on deep learning

Abstract

This article proposes a deep-learning-based fast coding unit (CU) partition algorithm to reduce encoding time of equirectangular projection (ERP) 360-degree videos in versatile video coding. First, an ERP 360-degree dataset with ERP latitude characteristics and quantitative parameter characteristics is established. Then, a lightweight prediction partition convolutional neural network is designed to predict the partition probability of 4$\times$4 CU edges in 32$\times$32 luminance CU. Finally, an intra prediction decision-making scheme is developed to reduce the number of candidate modes of CUs with size equal to or smaller than 32$\times$32, thereby achieving fast encoding. Experimental results show that the proposed method saves 58.51% of encoding time in All Intra configuration and only increases Bjontegaard delta bit-rate by 1.39%.