CNN-based ternary tree partition approach for VVC intra-QTMT coding

• Published in: Signal, image and video processing Vol. 18; no. 4; pp. 3587 - 3594
• Main Authors: Belghith, Fatma, Abdallah, Bouthaina, Ben Jdidia, Sonda, Ben Ayed, Mohamed Ali, Masmoudi, Nouri
• Format: Journal Article
• Language: English
• Published: London Springer London 01.06.2024; Springer Nature B.V
• Subjects: Artificial neural networks; Coders; Coding; Complexity; Computer Imaging; Computer Science; Decision trees; Efficiency; Image Processing and Computer Vision; Multimedia Information Systems; Original Paper; Pattern Recognition and Graphics; Signal,Image and Speech Processing; Video compression; Vision; Convolution Neural Network Ternary Tree (CNN-TT); Versatile Video Coding (VVC); Computational complexity; Quadtree with nested multi-type tree (QTMT)
• DOI: 10.1007/s11760-024-03023-5
• ISSN: 1863-1703

In July 2020, the Joint Video Experts Team has published the versatile video coding (VVC) standard. The VVC encoder enhances the coding efficiency compared with his predecessor high-efficiency video coding encoder, thanks to the improved coding modules and the new proposed techniques such as the new block partitioning structure called quadtree with nested multi-type tree (QTMT). However, QTMT induces a significant increase in encoding time mainly at the rate distortion optimization level (RDO) which causes an enormous computational complexity. Instead of RDO-QTMT partition process, a deep-QTMT partition approach based on a fast convolution neural network-ternary tree (CNN-TT) is proposed to predict the best intra-QTMT decision tree in order to reduce the encoding time. A database is initially established containing CU-based TT partition depths with several video contents. Then, a CNN-TT model is developed under three-levels provided by the TT structure to early determine the QTMT partition at 32 × 32. Different threshold values are fixed for each level according to the CNN-TT predicted probabilities to reach a balance between the encoding complexity and the coding efficiency. The experimental results prove that our deep-QTMT partition approach saves a significant encoder time on average between 23% and 58% with an acceptable RD performance.