Research On Data-driven Object Completion Algorithm

Image completion or inpainting aims at repairing damaged paintings or removingobjects in images. Traditional approaches are PDE-based diffusing or restoring the targetusing texture synthesis, which must sample in the original image. Problems remainedwhere users want to repair the target object in a picture. Object completion has twochallenges. Firstly, the target object might have its specific complex salient structure andtexture. Secondly, the missing parts of the target might be unable to sample in the picture.In this paper, we introduce a novel approach to image completion, which we calldata-driven object completion. In our system, the target object is completed through theprocess of structure completion and texture synthesis. In structure completion step, usershave to specify the shape of the target, and the system will retrieve the picture in the imagedatabase which has an object with the best matching contour and complete the targetstructure with samples from the best matching object. And the remaining missing parts arerestored by a target-guided texture synthesis method.Our system works well in tackling the two main issues of object completion. Contourmatching algorithm is first introduced into the study of image completion. And objectinformation is also taken into account to guide the texture synthesis. In experiments, oursystem produces better results than other state-of-art image completion method in repairingobject image. And with user control, our system can also yield different shape of the objectin the completion result with different shape specified in the structure completion step.