CU Partition Prediction Scheme for X265 Intra Coding Using Neural Networks

Abstract

This paper proposes an early termination algorithm that relies on a backpropagation neural network (BPNN) for predicting the decision of coding unit (CU) partition to avoid unnecessary computation and thus to accelerate HEVC intra coding process. One of the most important things for using BPNN is to discover suitable features that are profoundly correspondent to the way of making decision of CU partition and be helpful for training a model with high prediction accuracy. Block texture of CU is adopted as the input features for training data to model the HEVC behavior on CU partition. Experiment results show that it can decrease 40.78% average encoding time and increase only a little output encoded bitrate for most benchmark videos.