Study On Line Drawings-Guided Inpainting Method For Ancient Dunhuang Murals

Dunhuang murals,one of the major treasuries of Mogao Grottoes—a “World Cultural Heritage” site locates at northwest China.However,under the threat of many factors,murals have been haunted by various problems,which make them less informative and less beautiful.In recent years,manual repairs had been gradually replaced by digital inpainting techniques.Traditional methods often lead structure incoherence and texture inconsistence without incorporating structure information collected from line drawings,when repairing those serious damage murals.Especially in the filling step,traditional methods that use one single "special" match patch as best exemplar to inpaint damage areas,which leads continuous error-filling,like “Butterfly Effect” and “Block Effect”.To better address these problems,this paper conducts the following research.(1)This paper focus on the mural inpainting algorithm based on sparse representation,firstly we introduce the background and definition of sparse representation,and then present five sparse models under different norm minimizations used in sparsity constraints.Second,we categorized optimization algorithms based on sparse representation into four groups and summarize these algorithms.Finally,dictionary learning including the supervised dictionary learning and unsupervised dictionary learning are both introduced.Through the multi-angle overview on sparse representation theory,we sufficiently reveal the principle and potential properties of sparse models to supply important theory guidances for the study of Dunhuang mural inpainting algorthims.(2)We propose a line-drawings-guided inpainting algorithm for repairing the damage murals of Mogao Grottoes,Dunhuang to solve the structure incoherence and texture inconsistence.In the traditional inpainting algorithm,how to distinguish the structures and the textures is always a big problem,which often occur the unpleasant results.So we study the registration between damaged images and line-drawings to provide the guidance of structure for the proposed inpainting algorithms.The experiments show that the registration by human-computer interaction can complete the structure completion,and can also accurately locate the structures and textures in the damage area,which greatly reduce the rate of error-filling.(3)To better address the problems of patch selection and patch inpainting in traditional approaches,a novel patch selection scheme from texture patch to structure patch is designed.For texture patches located at fill-front,we first randomly select them without computing their priorities to improve the efficiency and texture coherence,and then define a patch structure complexity(PSC)and a new priority function to better determine the filling order for structure patches.Finally,we fill one patch with a combination of candidate patches by modeling the texture similarity and structure continuity in a sparse-representation framework.Experimental results demonstrate the effectiveness of the proposed method and avoid the “Butterfly Effect”.(4)To better address the problems of patch selection and patch inpainting in traditional approaches.We propose a global and local feature weighted method based on structure guidance to repair the damage murals of Yulin Grottoes and Mogao Grottoes,Gansu.Unlike traditional methods,we assume that damage regions can be estimated by global and local features,then we find that there has close connections between image inpainting and sparse representation,meanwhile,we analyze three different inpainting models including “lasso inpainting model”?“ridge inpainting model” and “elastic net inpainting model” in theory,so a novel sparse representation model with elastic net regularization based on similarity overcomplete dictionary is formulated to enhance the global feature consistency,and then an estimated method of neighborhood similarity is presented to guarantee local feature consistency,finally,we apply a global feature patch and local feature patch weighted method to obtain the target patch.Experimental results on damage murals demonstrate the proposed method outperforms state-of-the-art inpainting methods.