Initial-QP Prediction for Versatile Video Coding: A Multi-domain Feature-Driven Learning Approach.

Initial quantization parameter (Initial-QP) prediction is a key point in video compression, which directly affects the rationality of bit-rate allocation and the quality of subsequent encoded frames. However, the Initial-QP value is still determined based on some fixed parameters in the latest Versatile Video Coding (VVC) reference software, and it is difficult to reach the optimum. In this paper, we investigate the Initial-QP prediction on the VVC platform based on the Gaussian Process Regression (GPR) method. We explore the optimal relationship between Initial-QP and bit-rate under a variety of video content, and extract three different domain features to represent the complexity of video content. Moreover, we apply the proposed method for the I-frame rate control in VVC. Experimental results show that our method improves the accuracy by up to 7.69% with a slight computational complexity increase.