Research Of Image Inpainting Based On Geometric Information And Structural Feature

Image inpainting is the main research area in the domain of digital image processing. It can be used to conservation, remove redundant (or to restore old photo, remove text and hidden errors in videos, etc). In most case, it also has a great value in network transmission and damaged images.There are many approaches for image inpainting, then two kinds of important schemes in current methods, one is image inpainting based on the geometric imfomation models, which is according to partial differential equations (PDE); the other is image structural feature based on txture sythesis method. This thesis carries a deep study on those four parts through a great deal of experiments and acquires a series of valuable results which can be summarized in to the following aspects:First, on the basis of current Total Variation (TV) model, it used to repair small-scale cracks in the non-testure image. But it doesn't take full advantage of useful geometric information. Accodingly, it affects the inpainting speed and the edge of information. We proposed a adptive neighbored-weight of image inpainting based on TV model, The method uses a smooth decreasing function substitute of a fixed parameters. Quickly diffuse to damaged region in the early iterative process of repairing; maintain the detail of edge in the later.Second, the traditional Curvature-Driven Diffusions (CDD) model in which curvature is added can satisfy"connectivity principles", but the speed of the inpainting is relatively slow and the number of iterations is rather excessive in the corner. We proposed a fast diffusion image inpainting model by analyzing the effect of geometric information (gradient, curvature) in the iterative process of repairing. We only consider curvature in the initial iteration, the size of gradient will be used later. So it not only can recover large damage region, but also speed up inpainting time definitely.Third, this paper introduces the Criminisi algorithm in the texture synthesis; the current priority function will cause mutual inhibition of texture and structural feature. A new priortity function improved, it can optimize alogorthm and adjust order of the selected pixel block, so it can pay attention to the structural feature in the process of repairing.Fourth, a new scheme of image coding based on geometric information and structural feature. In image coding, extract image edge, then the large of smooth region is shape filter segmentation by mathematical morphology. Edge extension image and the part of target partition pixels are only encoded so that required information for encoding is greatly reduced. In decoding, a adptive TV model is applied to reconstrust the image.The experimental results show that with this scheme proposed, it can get a good quality of reconstructed image in the condition of less image information.