Deep Learning Based Low-Complexity andHigh-Speed Coding Unit Partition Decisionfor High Efficiency Video Coding

Abstract

High-Efficiency Video Coding (HEVC/H.265) is the current video coding standard and it achieves a 50% greater bitrate reduction than H.264.The Coding Tree Unit (CTU) is the main processing unit that definesthe HEVC decoding process. Due to recursive search method used inHEVC encoder to find the Coding Unit (CU) partitions, it increasescomputational complexity and time for processing. This paper suggeststwo approaches for coding unit partitioning, one is based on machinelearning and the other is based on deep learning. The parameterslike Edge Magnitude (EM), Rough Mode Decision cost (RMD), RateDistortion cost (RD), and Most Probable Modes (MPM) are used toestimate the complexity of CU splitting. These parameters are utilised bymachine learning methods, such as Fischer’s Linear Discriminant Analysis (FLDA), Support Vector Machine (SVM), and Logistic Regression(LR), to classify split and un-split samples of the coding unit partitioning. The machine learning algorithms increased the coding efficiencyand decreased its computing complexity in comparison to the traditionalHEVC encoder. However, machine learning algorithms also used the iterative procedure in determining the splitting of the coding units. Thispaper proposes an alternative method using deep learning technique toaddress the recursive search issue. Using a modified version of the LeNet-5 architecture, our proposed deep learning method simplifies coding unitpartitioning because neither recursive searching nor rate-distortion optimization are required to identify the coding unit partitioning in modifiedLeNet-5 architectures. As a result, the coding unit partition speed is raised in all intra-configurations without changing the computational costof HEVC. According to simulation results, the proposed techniques LR,SVM, and modified LeNet-5 could individually reduce nearly 50.09%,63.78%, and 86.06% of the time required to encode various test videos.