A fast video inpainting technique

Abstract

In this paper we present a fast video inpainting technique to infer the unknown information in the target region by maximizing box-based self-similarity and coherence measure. The video inpainting is already proposed in the literature and some of them are able to produce good quality results. However, the bottleneck of those algorithms is they are painfully slow. Here we fill the texture in the target region that preserves the smooth motion of the object without inclusion of any artifacts in reasonable amount of time. Our experiments show that the proposed method is quite efficient to synthesize unknown information in a video and comparable to the existing state-of-the-art methods. Moreover, proposed method is based on box filling and optimization is done on multiple scale using EM algorithm, and is computationally faster than the existing ones. © Springer-Verlag 2013.