In Versatile Video Coding (VVC), the partition patterns for coding units (CUs) have significant impact on the encoding efficiency. Determining the optimal CU partition is particularly time-consuming due to the calculation and comparison of rate-distortion costs for all possible partition patterns, especially during the ternary tree (TT) partitioning in intra coding. In this paper, a fast decision mechanism is proposed for TT partitioning based on image feature analysis to skip the complex rate-distortion calculation. Firstly, the correlation between the image structural features and the TT partition patterns is investigated based on experimental analysis and the most relevant features are selected for the subsequent prediction of optimal TT partition patterns. Secondly, we devise an efficient scheme for representing and extracting the selected features, further optimizing the extraction process to minimize computational complexity. Comprehensive datasets for partition pattern prediction are constructed based on these refined features. Finally, these datasets serve as the foundation for training and optimizing a predictive model, which is designed to achieve an optimal trade-off between prediction accuracy and model complexity. The predictive model is seamlessly incorporated into the VVC Test Model (VTM), facilitating efficient feature extraction prior to the Rate-Distortion Optimization (RDO) process for intra prediction and optimal partition pattern selection. By leveraging the prediction results, the model effectively determines whether TT partitioning can be bypassed, thereby streamlining the decision-making process and enhancing overall coding efficiency. Experimental results demonstrate that in comprehensive performance evaluations of time-saving metrics and Bjøntegaard Delta Bit Rate (BDBR), the proposed mechanism significantly outperforms existing lightweight neural network algorithms. Our decision mechanism effectively preserves coding quality while substantially accelerating the video coding process.
Sun et al. (Fri,) studied this question.