Removing Reflection From a Single Image With Ghosting Effect

Removing the undesired reflections of images taken through glass is an important problem in digital photography and many other vision applications. The so-called ghosting effect, i.e., the pattern repetitiveness in reflection, is an effective cue used by existing techniques to remove reflection from images. Existing methods take a two-stage approach that first estimates the parameters of ghosting effect and then models reflection removal as a two-layer separation problem: reflection layer and latent image layer. This paper aims at addressing one main challenge in such an approach, i.e., how to distinguish the repetitive patterns on the later image layer and the ghosting patterns on the reflection layer. Based on the observation that the number of repeats of natural image patterns is often different from that of ghosting patterns, we propose a wavelet transform based regularization method. Together with a novel weighting scheme, the proposed method is capable of accurately separating two layers, and experimental results justified its advantages over the existing ones on both synthetic and real data set.