Image Inpainting Method Based On FMM And Sparse Representation

Digital image inpainting is used for repairing and reconstructing the missing information of animage so that the viewer can not perceive the images have been broken or have been repaired at thescene. The current image inpainting methods can be divided into two categories: one is for smallmissing areas, and the other is for large missing areas. In this thesis, we focus on these two types ofimages to study a series of image inpainting algorithms and improvements, the main work andresults reflected in the following aspects:First, because of the linear and local feature of fast marching method(FMM), it is easy to beblurring on the repair boundaries and isophote, so we combine the gradient direction histogram toselect the direction of the surrounding pixels for the sake of the repair process in strict accordancewith isophote. Simulation results show that this method can maintain the edge and some rulestexture better and the repair effect has been greatly improved.Second, as the image inpainting method based on a sample sparse representation can not keepthe linear structure well when repairing the image with large defect of structural information, wepropose a method by structural constraints and sample sparse representation. The image edgeinformation is repaired by a polynomial curve fitting to constrain the structural information. Then anarrow-band model of sample sparse representation is used to repair structural information. Thetexture information is completed by a translational block sparse representation method. Simulationresults show that proposed method can achieve higher image quality, can better repair the structureinformation and maintain the smoothness of structure integrally.Third, as sample image dictionary has poor adaptability and the unity of valid information,which resulting in bad images sparse representation, we propose a new image inpainting method bycharacteristics classification learning and patch sparsity propagation. This method classified theimage patches by their different characteristics firstly, then get the corresponding over-completedictionary by training the image blocks of different characteristics and extracted different validinformation from these blocks for sparse coding. Finally, we modified the patch sparsitypropagation model to improve the propagation mechanisms. Simulation results show that proposedmethod can works well on the edge, irregular textures and smooth portion and make higher imagequality.