Spatial error concealment via model based coupled sparse representation

In this paper, we propose a novel spatial error concealment algorithm through model-based coupled sparse representation. According to the non-local self-similarity property of natural images, we first collect two set of samples by template matching: one is called the latent set corresponding to the current missing patch and the other one is called the template set corresponding to the current template. Using these two sets of samples as the training data, we learn a dictionary pair and a linear prediction model simultaneously. The pair of dictionaries aims to characterize the two structural domains of the two sets, and the linear model is to reveal the intrinsic relationship between the sparse representations of the current missing patches and its template. Finally, we cast the non-local dictionary learning and local correlation model into a unified coupled sparse coding framework to obtain optimal sparse representation and further accurate estimation of the current missing patch. Experimental results demonstrate that the proposed method remarkably outperforms previous approaches.