CtuNet: A Deep Learning-based Framework for Fast CTU Partitioning of H265/HEVC Intra- coding

Abstract

Nowadays, real-time multimedia applications mandate high video quality while maintaining reasonable bitrates. The H.264 coding delivered inexpensive bitrate costs compared to other coding schemes while maintaining high-grade video quality, yet bounded to deliver higher qualities. Later, High-Efficiency Video Coding (HEVC) improved on H.264 by providing higher video qualities with an efficient bitrate. However, such improvement obligates higher computational expenses due to employing superior techniques like quad-tree for coding tree unit (CTU) partitioning. This paper proposes a framework, named CtuNet, for CTU partitioning by approximating its functionality using deep learning techniques. A ResNet18-CNN model is adopted to predict the CTU partition of the HEVC standard. We have baselined our suggestion with state-of-the-art approaches. The results demonstrate the supremacy of the proposed CtuNet over the other approaches. The CtuNet framework maintains near-optimal results by reducing computational complexity up to 63.68% with negligible degradation in bitrate by 1.77% at intra-prediction.

Keywords

• Deep learning
• Convolutional neural network (CNN)
• ResNet
• H265/HEVC
• Intra-prediction
• Video coding