A Robust Image Reflection Separation Method Based on Sift-Edge Flow

摘要

Motion computation is an important part for image reflection separation methods using image sequence. As the motion of background and reflection is separated on image edges, in this paper we present a robust image reflection separation method based on sift-edge flow. Edge motion is firstly computed using sift-edge flow. Based on it, the sparse motion field of image edges is clustered into two parts and refined by RANSAC, which are then used to interpolate the dense motion fields of the background and reflection. The initial image decomposition is conducted by warping the input images to reference image using the dense motion fields of background and reflection respectively. Finally, with the initial solution provided, the background and reflection images can be separated using alternating optimization. Experiment results showed that our method can achieve a robust performance compared with the state of art.