Research On Exemplar-Based Digital Image Inpainting

Exemplar-based image inpainting is the classical technique of texture-based image inpaintingtechnique, but currently it has the problems of low efficiency and poor quality in inpainting. Tosolve the problems, this thesis discusses the improved methods of image inpainting from the localsearching of best-matching, updating confidence values, the inpainting order and selecting the sizeof inpainting patch adaptively to improve image’s quality and speed. The main works andconclusions are described as follows.1. This thesis first systematically introduces the background, goal and meaning of digitalimage inpainting technique, and summarizes its development state of home and abroad, broadapplication, mathematic analysis and objective assessment of inpainting results. Then this thesisstudies the basic application principle and flow of some classical texture-based image inpaintingalgorithms and analyzes their strong and weak points.2. For the problems of low efficiency in inpainting process and error accumulation easily inCriminisi’s algorithm, based on the Markov Random Field theory, this thesis presents a method todetermine the local searching area, in which all candidates are included, according to the damagedarea in the way of searching area of best-patch and confidence updating. A number of experimentsshow that proposed algorithm not only largely reduces the inpainting time, the inpainting quality isalso improved to a certain degree.3. Add the pixels in corresponding position at best-matching patch and target patch satisfy thepixel similarity as additional matching condition; meanwhile, according to the difference values ofknown pixels in the neighboring area of a damaged pixel, decide the type of boundary: strongboundary, comparatively strong boundary or weak boundary. Under the new filling order, this paperprovides a method of selecting adaptively the size of matching patch on the basis of the boundarystrong or weak, in the damaged pixel’s neighboring area. Simulation results demonstrate thatproposed algorithm improve the image quality to great degree, and the inpainting results satisfy thehuman visual system better.