Enhanced Machine Learning-based Inter Coding for VVC

Martin Benjak, Holger Meuel, Thorsten Laude, Jörn Ostermann

Abstract: In this paper, we propose an enhanced machine learning-based inter coding algorithm for VVC. Conceptually, the reference pictures from the decoded picture butter are processed using a recurrent neural network to generate an artificial reference picture at the time instance of the currently coded picture. The network is trained using a SATD cost function to minimize the bit rate cost for the prediction error rather than the pixel-wise difference. By this we achieved average weighted BD-rate gains of 0.94%. The coding time increased about 5% for the encoder and 300% for the decoder due to the use of a neural network.