Image inpainting based on an improved K-SVD counterpart algorithm

An improved K-SVD counterpart algorithm for image inpainting with a new strategy for patch extracting is proposed. In the new framework, firstly a two layers patch extracting approach is proposed to enable the process of image inpainting on each patch more effectively. Then the K-SVD counterpart dictionary learning based on each corrupted patch is realized with alternating minimization strategy. Each patch is inpainted through the framework, which will be integrated and averaged to form the final visible image. Simulations on several popular used images in different missing cases illustrate that the proposed algorithm improves the recovering performance in term of peak signal-to-noise ratio (PSNR).