VMAF Oriented End-to-End Quality Prediction Model for Perceptual Versatile Video Coding

Recently, Netflix has developed VMAF (Video Multi-method Assessment Fusion) model, a comprehensive video quality assessment (VQA) model that combines spatial and temporal characteristics to predict perceptual quality for video. This model has shown stronger correlation with the human visual system than traditional VQA metrics like PSNR (Peak Signal to Noise Ratio) and SSIM (Structure Similarity Index). However, perceptual video quality often degrades due to the quality and bitrate variation in reconstructed videos, leading to uncomfortable viewing experiences. In this context, we propose a novel quality prediction model for perceptual versatile video coding (VVC). The proposed quality prediction model compromises a VMAF measurement and an end-to-end machine learning based quality prediction algorithm. The predicted perceptual quality is then fed into a Rate-Distortion (RD) optimization process of VVC video encoding to create the quantization parameters map which is finally used to lossy compress the video. Experimental results revealed that the proposed perceptual VVC solution is able to achieve much lower the quality fluctuation and meet the user quality requirement while asking fewer compression bitrate than the standard VVC.