Automatic Image Inpainting by Heuristic Texture and Structure Completion

This paper studies an image inpainting solution based on a primal sketch representation model [1]; which divides an image into structure (sketchable) and texture (non-sketchable) components. This solution first predicts the missing structures, such as curves and corners; using a tensor voting algorithm [2]. Then the texture parts along structural sketches are synthesized with the patches sampled from the known regions, and the remaining texture parts are defused using a graph cuts algorithm [3]. Compared to the state-of-art image inpainting approaches, the characteristics of this solution include: 1) using the primal sketch representation model to guide completion for visual consistency; 2) achieving fully automatic completion. Finally, the experiments on the public datasets show above characteristics.