VGG-19-based parameter optimisation for R-lambda model for VVC intra rate control
Huiyuan Sun, Xuesong Jin · 2023
Intra rate control (RC) is very important in the whole video coding, in which obtaining the actual optimal initial model parameters for each intra frame is a challenge since there is no a priori information that can be relied upon for each intra frame and the R-lambda model parameters α, β of each coding tree unit (CTU) of an intra frame are kept constant and are not updated with the coding CTU. In the intra frame rate control of Versatile Video Coding (VVC) standard, this paper proposes an optimal α, β decision based on Visual Geometry Group 19(VGG-19). Firstly, a convolutional neural network (CNN) dataset is built based on a natural image dataset with the default bit rate control model of H.266/VVC standard. Second, the basic VGG-19 model is improved to adapt the self-built training set and predict the key parameters α, β of the R-lambda model. Finally, different models are trained by the dividing line of α in the dataset, and the one that can make the most improvement in the efficiency of the rate control is selected to embed the encoder. The experimental results show that the algorithm can reduce the average error of the bit rate control up to to 4.67%.