Research And Application On Digital Image Inpainting

Image inpainting is an important research topic in the area of image restoration. Its objective is to restore the missing or damaged portions of the image in order to make it more legible and to restore its unity in a way that is non-detectable for an observer who does not know the original image. Currently, digital inpainting techniques have found broad applications in image processing, vision analysis, digital restoration of ancient paintings for conservation purposes, text removal and objects removal in images for special effects, restoration of old photographs or films with scratches or missing patches, disocclusion in computer vision, errors conceal in videos, and so on. However, everything is just at the beginning. Based on understanding and utilization of the previous achievements, this dissertation focuses on developing innovative digital image inpainting methods.This dissertation attempts to research on image inpainting techniques and image completion techniques, as well as their applications. The former is suitable to inpaint the small scale scratches in images while the latter is very good at completing the large objects. Firstly, it researches on the variational PDE image models and corresponding algorithms for image inpainting which is the research hotspot recently. Then, it focuses on image completion based on texture synthesis to fill in the large damaged objects. Finally, it discusses the broad applications of image inpainting including image compression, image zooming and digital restoration of ancient paintings.Chapter 1 is the preface of this dissertation, which introduces the background knowledge, reviews the state of arts development and broad applications of digital image inpainting techniques, and summarizes the central research work and innovative points in this dissertation.As a theoretical guide for practice work, Chapter 2 introduces some review on the basic principles related to image inpainting including the "best guess" principle and Bayesian framework, variational basic theory, total variation inpainting model as well as texture synthesis.Chapter 3 studies image inpainting technique. It is presented that two variational inpainting models based on p-harmonic energy minimization depending on whether or not noise needs to be suppressed in the image. Deducing the associated Euler-Lagrange equations of the two models, it is analyzed the diffusion performance of the p-harmonic equations and proved the two models have unique solutions in the space of Sobolev. Finally, with half-point differential scheme, a finite difference inpainting algorithm is proposed. Theoretic analysis and experimental results show that the new models have better performances both on vision effect and convergence speed than total variation inpainting model.Chapter 4 studies wavelet image inpainting technique. A wavelet inpainting model based on p-Laplace operator for the lost wavelet coefficients is presented. This new model is suitable for processing noisy images and noise free images through adjusting the alterable model parameters. It then deduces the associated Euler-Lagrange equation and established the corresponding diffusion equation. Finally, a finite difference inpainting algorithm is proposed. Theoretic analysis and experimental results show that better inpaingting quality can be achieved with much less computing time with the new model.Chapter 5 studies image completion techniques. It presents a fast adaptive algorithm based on texture synthesis for natural image completion. The algorithm takes advantage of the fact that most natural images have some directional distribution of texture and color. So the search is made over a smaller region to reduce the computational complexity. It defines the searching order of the patches to ensure the regions with more known information and structures should be completed before filling in other regions and presents an adaptive scheme to determine the size of template window to improve the quality of the output image. Comparative experiments show that the proposed algorithm outperforms the earlier works in terms of both perceptual quality and computational efficiency.Chapter 6 discusses the broad applications of image inpainting including image compression and zooming. Firstly, in image compression, image inpainting method is applied to image compression and a novel scheme of compression based on edge information is proposed. In encoding, it is only encoding edge extension image so that creating a large area of blank domains where the image information has been wiped out, and thus a high compression rate is achieved. In decoding, an image inpainting method is applied to reconstruct the image. Experiment results show that with this scheme proposed, one can get a good quality of reconstructed image in less image information. Secondly, in image zooming, we establish a digital p-harmonic filter based on p-harmonic image inpainting model to noisy image zoom-in. Experimental results show that the digital p-harmonic filter achieves better performance, it can zoom an image with arbitrary accuracy and denoise.Chapter 7 summarizes the dissertation and brings forward some problems that need further researching on.