Image Inpainting With Data-Adaptive Sparsity

Image inpainting finds numerous applications in object removal, error concealment, view synthesis, and so on. Among the existing methods, exemplar-based inpainting has been shown to achieve superior performance when filling in large areas. In this work we study inpainting based on sparse representations, as a generalization of conventional exemplar-based inpainting. The particular novelty is the data-driven adaptation of the sparsity level according to the strength of linear structures incident on the fill front. Experimental results show that the proposed method achieves improvement in both subjective and objective inpainting performance compared to well-known exemplar-based inpainting.