Image error concealment based on joint sparse representation and non-local similarity

Abstract

In this paper, an image error concealment method based on joint local sparse representation and non-local similarity is proposed. The proposed method obtains an optimal sparse representation of an image patch, including missing pixels and known neighboring pixels for recovery purpose. At first, a pair of dictionary and a mapping function are simultaneously learned offline from a training data set. The dictionary pair transfers the original image patch and corrupted image patch into a domain that the intrinsic relationship between patches can be derived more accurate than the spatial domain. Then, in the error recovery process, the sparse representation of corrupted patch is obtained and mapped using the mapping function. The obtained sparse representation is further refined by exploiting the self-similarities in the natural images to obtain a good estimate of the sparse representation on the original and clean image. Finally, the concealed image is obtained by back projection of this sparse presentation into the spatial domain. Compared with the state-of-the-arts error concealment algorithms, experimental results show that the proposed method has better reconstruction performance in terms of objective and subjective evaluations in different loss scenarios and validate its effectiveness for the error concealment applications.