Deep Learning-based Transformation Matrix Estimation for Bidirectional Interframe Prediction
Satoru Jimbo, Ji Wang, Yoshiyuki Yashima · 2018
In this paper, we propose a new method to apply deep learning to bidirectional interframe prediction in video compression. The novelty of the proposed method is to create an interpolated frame by the geometric transformation matrices estimated by CNN whose inputs are temporally previous and future frames. The proposed method can achieve considerably higher efficiency for bidirectional prediction because the geometric transformation matrix estimated by learning can express parallel translation, zoom in/out and change of blurriness with arbitrary accuracy. Experimental results show the prediction error reduction of over 30% compared with H.265/HEVC, especially for video sequences with small motion.