Automatically Integrated Learning Technology-based Digital Image Repair Theory And Algorithm Research

Digital image inpainting techniques is an important research topic in the area of image restoration, it is a process of filling in the missing data in a damaged region of an image in a visually plausible way, and which can get the image to the extent that the inpainted image looks continuous, intact and natural perceptually. Currently, image inpainting have found broad applications in image processing, digital restoration of ancient paintings for conservations purposes, restoration of the old photos, text removal and objects removal in images for special effects, vision analysis and so on. Ensemble learning is a new machine learning paradigm; it can significantly improve the generalization ability of learning systems through utilizing multiple learners to solve a problem. This paper attempts to research on digital image inpainting techniques base on the ensemble learning techniques. The main content of this dissertation is described as follows.(1)The background, purpose and applications of the digital image inpainting technology are introduced in the paper, and the mathematical theories of inpainting problems are also discussed briefly. Several classic inpainting models and their inpaingting principles are overviewed in details, such as BSCS, TV, CDD, Mumford–shah and exemplar-based image inpainting method and so on. The merit and demerit of the above algorithms are analyzed through the inpainted images.(2)A new algorithm of image inpainting based on Ensemble learning technology, namely an improved object removal inpainting method is presented in the paper. This algorithm is an object removal method based on exemplar inpainting and can obtain improved performance over exemplar-based inpainting methods. The formula of priority block in Criminisi's method is optimized and the search precision is improved. The best match patch is found by using the ensemble learning technology to learn the sample characteristics, and the algorithm runs faster. Experimental results on nature images show that the improved method can greatly accelerate the inpainting speed and get the better results.(3)A new image inpainting approach is proposed by solving a Poisson equation and integrating gradient information. First, the best match patch is found by using the gradient information. And then, the image is reconstructed from the gradient maps by solving a Poisson equation. The experimental results show the method works efficiently in large area inpainting tasks.(4)A concept of blind image inpainting is presented. It base on exemplar-based image inpainting and improves the image inpainting algorithm base on texture synthesis. Trustworthiness judgment model of the digital image is used in the proposed image inpainting method to get the areas of the image which need to be repaired. Experiments show that the algorithm can obtain the effect of the automatic blind image inpainting.