Intra Prediction Mode using Conditional-UNet by Stochastic Differential Equations for Next Generation Video Coding

With the spread of high-resolution video such as 8K and 360-degree videos, efficient video coding methods is still highly required. The intra prediction method of versatile video coding (VVC) achieves efficient coding by referencing neighboring pixel values based on predefined directional modes. However, it is difficult to further improve video coding efficiency using traditional linear models in the intra prediction mode. In this work, we propose a method to increase the coding efficiency by applying the image restoration technique. Specifically, image restoration technique uses Conditional-UNet combined with a stochastic differential equation (SDE) to generate highly accurate predicted pixels. The simulation results show that the proposed algorithm can achieve an improvement of 29.14 % on BD-rate compared to the original VVC algorithm at random access. Moreover, the proposed algorithm reduces the training and inference time through optimized parameters.