Frame Rate Up-Conversion Using Bi-Directional Optical Flows With Dual Regularization

摘要

In this paper, we present a frame rate up-conversion (FRUC) scheme that uses optical flows in forward and backward directions and with two different regularization parameters to derive alternative motion vector fields. The motion vectors are then assessed for their relevance, resulting in prediction weights assigned to each candidate motion vector. Multiple hypothesis reconstruction is then performed, where overlapping patches of pixels associated with multiple motion vectors are fused using their corresponding weights to yield pixels of an interpolated frame. Four flavors of our approach are proposed having different tradeoffs between quality and computational complexity. Experiments conducted using standard test set show that all our schemes yield significantly higher quality compared to existing FRUC algorithms.