A Research And Application Of Exemplar Structure Based Image Inpainting

Digital image inpainting is currently a major research hotspot in the field of image processing and computer vision. It uses the information of the known region to fill in the damaged region of the image by certain criteria, such as the known region and the damaged region share the similar geometry or statistical characteristics, so that the retrieved image achieves the desired visual quality and people will not notice the trace of the inpainting. This technology has a wide range of application scenarios. It plays an important role in ancient artifacts repair, target object removal, image lossy compression and high-resolution studies.Diffusion-based inpainting and exemplar-based inpainting are two main categories of current leading image restoration methods. Compared to diffusion-based inpainting methods, which are applicable to images with small damaged regions of structural information, exemplar-based inpainting methods restore the images using statistical properties and self-similarity of the image texture information, and are applicable to natural images and images with large damaged regions. However, these methods have their own flaws of unreasonable restore order, inaccurate matching criteria, highly time-consuming searching process and so on. This paper conducts in-depth research on image inpainting methods and proposes an exemplar-based inpainting using structural consistent patch matching and an exemplar deformation based inpainting.(1) Exemplar-based inpainting using structural consistent patch matching. In this paper, we propose a space varying updating strategy for the confidence term, eliminating the problem of improper filling order caused by dropping effect of the confidence term. Furthermore, we introduce a matching confidence term to improve the filling priority estimation based on the matching quality of filled pixels. This paper presents a structural consistent patch matching criteria to eliminate the mismatch caused by the original inaccurate matching criteria and converts the search of the best match to the frequency domain using FFT to improve the efficiency of the algorithm.(2) Exemplar deformation based inpainting. After imposing the previously proposed algorithm on images, there is still some certain incoherence in the texture region and dis-connectivity in the structural region. To solve this problem, this paper conducts sub-pixel registration between the best match patch and the patch to be restored, maximize the similarity between these two patches and by then restore the image.The algorithms proposed in this paper are able to restore images with large damaged regions filled with both structural information and texture information and retain its information. Through various experiments of damaged images, restored image results retrieved by methods proposed in this paper are more preferred to human visual habit compared to the traditional inpainting algorithms.In addition, this paper develops an image inpainting prototype system based on B/S architecture using ActiveX and implements the proposed algorithms in it. The system provides operations, such as importing images, determination of the damaged region, selection of parameters and so on.