Image Inpainting Algorithm Based On Exemplar Structure Characteristics

With the continuous development of computer and multimedia technology, the digital image inpainting technology is applied more and more widely, has been the concern and attention of more people. In this thesis, firstly the status of digital image inpainting technology is introduced based on different classifications, and typical models and algorithms are described simply. The Criminisi algorithm of the typical exemplar-based algorithms and the comparative literature algorithms is analyzed and described in detail, meanwhile the inadequacies in the current exemplar-based image inpainting algorithms are summed up, this thesis deep studies the problems of the exemplar-based inpainting algorithms.In the current exemplar-based image inpainting algorithms, there are some shortcomings such as the unreliability of the filling order, the improper template window of sample block and the low efficiency of inpainting etc. To address these problems, an image inpainting algorithm is designed based on exemplar local structure properties. Combining with the local gradient features and characteristics of local similarity of image, the proposed algorithm increases the compensation coefficient and weight of the data term to ensure the connectivity of image structure. The size of template window of sample block is chosen by the gradient information of image to reduce error derivative and block effects phenomenon. The confidence term is used to adaptively select the size of local search space to improve efficiency of inpainting. Experimental results show that the proposed algorithm can maintain the structure characteristic of the damaged region and reduce error derivative and block effects phenomenon.When repairing damaged image with image inpainting algorithm based on exemplar local structure properties, there are still some problems such as the jagged effect in texture information-rich region, the poor connectivity of the edge of structure and mismatch phenomenon. To address these problems, this paper designs an image inpainting algorithm with matching criterion based on image direction characteristic obtained by the multi-scale and multi-directional characteristics of the Curvelet transform. The eight direction factors of the pre-repaired image are estimated by using the Curvelet transform, and then new matching criterion is constructed based on the weighted distance of direction factor and color between the target block and the matching block. The weight coefficient and effectiveness of the matching criterion and the impact of the pre-repair method to inpainting result are analyzed and discussed. Experimental results show that the improved matching criterion based on direction factor can preferably preserve the texture and structure properties of the image; compared to the existing similar algorithms, the quality of the texture block and structure block repaired by the proposed algorithm increases about0.8dB and1.2dB respectively.In order to demonstrate easily the image inpainting algorithm of this article, enhance its usefulness, this thesis develops a digital image inpainting demonstration system based on Visual Studo2008and MATLAB platform, at the same time, provides the instruction and demonstration results of the system.