Self-Similarity Inpainting

We present a novel means of texture inpainting, dubbed self-similarity inpainting, that uses the self-similarity descriptor to implicitly encode complex structures through a summary of local neighborhood comparisons. Pixel patches are selected from a codebook based on descriptor distance in their original locale and after the proposed insertion. We suggest an efficient means of parallelizing this approach across an arbitrarily large number of processors, as well as describing improvements over existing techniques and extensions that shift the tradeoff between inpainting quality and algorithmic efficiency for object or artifact removal. Results are shown for a number of synthetic and captured digital images, including effects upon human foveal attention.