Deep Learning-based Transformation Matrix Prediction for Video Frame Interpolation

Abstract

In this paper, we newly propose a frame interpolation method using transformation matrices generated by deep neu ral network. The matrices can deal with many geometric transformations such as arbitrary precision motion /zoom compensation, luminance compensation and blur compensation. Experimental results show that the proposed CNN - based interpolation method gives better performance for video seque nces with small motion than the conventional HEVC - based one .