A Multiscale Method for Automated Inpainting

We present a novel, simple and general method for image inpainting. Current methods may be crudely divided into those that aim to continue edges by various energy minimisation techniques, and those that perform texture synthesis from local information, but both have their weaknesses. Our method searches the image for areas of similarity and uses these to inpaint. By analysing the image at multiple resolution scales we can find similar features and textures from anywhere in the image at a reasonable speed. We present results using some challenging images where both features (edges) and textures from non-local information are used to achieve plausible restoration.