Speed Up VVC Intra-Coding by Learned Models and Feature Statistics

Abstract

The newest video coding standard, Versatile Video Coding (VVC), adopts a QTMT, quad-tree plus multi-type tree (MTT), block partition structure and improves the compression performance by about 30% ${\sim }50$ %, compared with the previous High-Efficiency Video Coding (HEVC) one, at the cost of higher time complexity. To make practical video communication applications feasible, we have to reduce the high time complexity resulting from an exhaustive rate-distortion optimization (RDO) search procedure. We proposed to predict Coding Unit (CU) split modes by a learned model whose input comprises neighboring line pixels and quantization parameters (QP). In addition, we set thresholds based on statistical image features and coding behaviors to eliminate unnecessary coding operations in critical coding control modules. Experiments showed that, compared with the default VVC intra-coding, the proposed method saves 46.73% of encoding time, with Bjøntegaard Delta Bit Rate (BDBR) increment of 1.16%. After retraining the learned model with a specified QP, the time reduction rate can achieve 51.79%, and the BDBR slightly increases to 2.07%. The proposed speedup coding scheme effectively reduced the VVC time complexity to a large extent.

Keywords

• Versatile video coding (H.266/VVC)
• coding unit split mode
• convolutional neural networks
• intra-coding acceleration