Image Inpainting Research Based On Sparse Representation With SL0

Digital image inpainting is a process that utilizes a certain algorithm to reconstruct the damaged area or eliminate the redundant parts in image,which make the restoration of the image be more in accordance with visual effect,and human eyes can't observe the repair scars.The image inpainting method based on sparse representation has become more valuable.Because it overcomes the problem of error matching in traditional image inpainting method based on texture synthesis,and it has more adaptive ability.In this thesis,based on the sparse representation inpainting algorithms,it analyzes their shortcomings and explores a series of improvement measures.The main work and achievements are follows:First,in the traditional image inpainting algorithm based on sparse representation,the reconstruction algorithm always makes use of the orthogonal matching pursuits(OMP)algorithm.However,it needs to estimate the sparsity of the image block,and the sparse degree of all image blocks are set to a uniform value,while no matter the sparse degree estimation is too small or too large will affect the reconstruction results.This thesis proposes a novel image inpainting algorithm by classified sparse representation based on the smoothed 0l norm(SL0)algorithm.The proposed algorithm firstly divided the damaged image into blocks,according to image blocks' features to classify them;then train the image blocks sample with the singular value decomposition method(SVD)to get the corresponding over complete dictionary to improve the adaptive of dictionary.Based on traditional SL0 algorithm,it takes use of approximate hyperbolic tangent function to approach 0l norm,in addition,using conjugate gradient to solve this function,which are helpful to improve the reconstructive accuracy.Experiment results show that the image inpainting quality of the proposed algorithm is greatly improved,which is more in accordance with visual effect.Second,the traditional sparse representation inpainting algorithms will make edges fuzzy and bring extension when fixes the edges.To overcome these shortages,this thesis utilizes edge fitting technology and edge structure constraints to propose a new image inpainting method by SL0 with edge fitting.Proposed algorithm firstly extracts edges integrally and extracts the edges of damaged area locally,then obtains kinds of information of damaged edges to repair the broken edges,which ensure edge structure is strictly restricted by the edge contour.Experiment results show that the proposed algorithm preserve the coherence of the image edge,and overcome the phenomenon fuzzy in edges,the image inpainting quality of the proposed algorithm has better visual effect.