Research Of Image Inpainting Algorithm Based On Texture Synthesis

Image inpainting is an important topic in the field of digital image processing, it belongs to the area of image restoration. Image inpainting is a technique to estimate and inpaint the damaged parts of an image using the rest image information. The aim of this technique is to remove a target object in an image and restore the missing or damaged regions in a visually plausible way that an ordinary observer cannot detect the flaw. The technology of image inpainting has broad application prospects, such as old photograph scratches recovery, image editing, virtual reality and transmission for network data.There are mainly two kinds of image inpainting techniques。Most of early works focus on filling in small image gaps using the theory of partial differential equation(PDE). These methods diffuse the already known image information into unknown regions in pixel level. The diffusion-based inapinting methods generate smooth effect when coping with large damaged regions or textured parts. There is also another category to restore images with large damaged parts basing on texture synthesis. These techniques use the already known image information to synthesis the to-be-filled regions in patch level. Criminisi designed an isophote-driven image sampling process and the restoration results are within the state-of-the-art.Inspired by the work of Criminisi and to resolve its problems, a new confidence updating rule is raised. The confidence values of those patches that filled better are set relatively high. In addition, the color variations of adjacent pixels are introduced into the priority function. Besides, gradient distance is utilized to determine the best match patch when there are several most similar blocks. Details of the proposed algorithm are as follows:(1) Criminisi’s algorithm generates a poor result if the image has strong structures damaged. In order to solve this problem, color variations term is introduced into this work. It is more likely that the opted block contains strong structures if the color variations term has a high value. Therefore it should be completed first. It is obvious that the image structures yield high priorities due to using the color variations term.(2) In conventional image inpainting algorithm based on texture synthesis, the best source patch is searched by calculating the sum of squared difference of known pixel’s colors. Considering there is much similar and repeated information in a practical image, a target patch may have several similar blocks with a same minimum color distance. In that case, the final best patch is decided randomly. Therefore, we utilize the gradient information to distinguish these blocks with a same color distance. It is known that gradient is an important geometrical feature of an image. It represents the image structures in a certain extent.(3) Filling order is critical in the exemplar-based inpainting algorithm. If an image block was filled, the pixel information becomes known and its confidence value should be updated. The updating of confidence term in Criminis’s algorithm does not consider the quality of exemplars filling. The credibility of filling order decreases as the completion carrying out. To obtain a reasonable priority, the similarity of the target patch and the source patch is introduced into the updating course. Using the improved updating rule, the patches that filled properly will have higher priorities and be synthesized first.