Image completion with patch sparsity-based global optimization

This paper presents a novel approach for image completion using global optimization, which combines with patch sparsity-based priority and dynamic structural label pruning. In our approach, the completion problem is described by discrete Markov Random Field model with a well defined objective function, and can be solved by adopting belief propagation. Two important extensions are proposed in the paper: 1) Priority based on patch sparsity. The scheme provides a reasonable synthesizing order for belief propagation algorithm, which encourages regions with salient structures to synthesize first. 2) Dynamic structural label pruning. To restrict faulty label candidates, we add structural information into previous label pruning step. We demonstrate our approach can complete natural images and photographs coherently.