Deep Learning For Intra Frame Coding
Amna Maraoui, Imen Werda, Sayadi Fatma Ezahra · 2021 International Conference on Engineering and Emerging Technologies (ICEET) · 2021
The evolution of the coding unit module from the High Efficiency Video Coding (HEVC) video standards to the Joint Exploration Model (JEM) extensively enhanced compression performance while severely increases coding complexity caused by the brute force search built on Rate Distortion Optimization (RDO). Effectively, compared to the predecessor HEVC standard that makes use of the quad-tree (QT) block partitioning module, the novel quad-tree binary-tree (QTBT) block partitioning process proposed within the JEM encoder uses additional block sizes and shapes which induces additional flexibility. In this paper, we suggest a deep convolutional neural network (CNN) based method to reduce the block partitioning module complexity for both HEVC and JEM encoders at all intra-configurations. The results show that the CNN approach leads to better optimization performance with the HEVC encoder reaching up to 59%. However, the CNN model is more robust with several JEM versions.