Research On Dunhuang Mural Inpainting Based On Texture Synthesis And Deep Learning

Murals are the crystallization of the wisdom and art of the ancient people.They not only record the living customs,religious beliefs and artistic culture of the ancient people,but also carry the thoughts and connotations of the Chinese nation for thousands of years.The Mogao Grottoes in Dunhuang are the largest art sanctuary in the world.The murals and scriptures contained in them have precious research value in many fields such as history,religion,ethnicity,and art.However,due to the long history of excavation of the Mogao Grottoes and the harsh geographical environment in which they are located,most of the murals in the caves have fallen off,cracked,mildewed,faded and other problems,which not only limited the development of Dunhuang cultural research,but also affected the ancient times.Cultural conservation research has taken a toll.Therefore,the research work on the inpainting of Dunhuang murals is imminent.At present,the inpainting method for diseased murals is mainly that researchers use professional knowledge to directly inpainting the diseased murals.However,this inpainting method is time-consuming,costly and irreversible,which poses a potential threat to the inpainting of murals.Therefore,using digital methods to research and inpainting virtual murals can not only display the inpainted murals virtually,but also provide an important reference for the inpainting of artificial murals,avoiding problems such as irreversibility in the process of mural inpainting.This thesis analyzes the research status of image inpainting at home and abroad,and on this basis,deeply explores the related algorithms of digital image inpainting.The main contents include the following three aspects: Dunhuang mural inpainting algorithm based on sequential similarity and cuckoo optimization,missing Edge reconstruction and priority-improved Dunhuang mural inpainting algorithm and multi-scale feature and attention fusion Dunhuang mural inpainting algorithm.The main research and contributions are as follows:(1)Aiming at the problems of error inpainting,long time and low efficiency in inpainting mural images by texture synthesis image inpainting algorithm,a Dunhuang murals inpainting algorithm based on sequential similarity and cuckoo optimization is proposed.First,the p-Laplace operator is used to redefine the data items to improve the priority calculation method.Secondly,the Sequential Similarity Detection Algorithm based on dynamic thresholds is introduced to find matching blocks to improve the inpainting efficiency of the algorithm.Finally,in order to improve the accuracy of the inpainting results,the cuckoo optimization algorithm is introduced to optimize the matching blocks.Through inpainting experiments on damaged Dunhuang murals and comparing with the texture synthesis image inpainting algorithm,the experimental results show that the proposed algorithm can effectively inpainting the damaged murals,and the inpainting efficiency has been further improved.(2)In view of the complex line structure of damaged Dunhuang murals,the traditional image inpainting algorithm based on block matching cannot properly inpainting the missing structures when inpainting the murals,a Dunhuang mural inpainting algorithm with edge missing reconstruction and priority improvement is proposed.First,the missing edge structure in the damaged mural image is reconstructed by the Bezier curve fitting method to enhance the structure of the mural.Secondly,the gradient,curvature and other characteristic factors are introduced to improve the priority calculation method,making the priority calculation more reasonable,to avoid the problem of wrong filling caused by the priority tending to 0.Finally,the proposed algorithm and the comparison algorithm are used to inpainting the damaged Dunhuang murals.The experimental results show that the proposed algorithm can inpainting the damaged mural structure better than the comparison algorithm,and the peak signal-to-noise ratio and structural similarity have been improved.(3)Aiming at the problems of low feature utilization and lack of image texture details after inpainting damaged murals by existing deep learning image inpainting algorithms,a multi-scale feature and attention fusion network model for Dunhuang mural inpainting is proposed.First,a multi-scale feature fusion generator is designed to extract and fuse the feature information of different scales of murals to improve feature utilization and enable the network to learn more comprehensive image features.Secondly,the self-attention mechanism is introduced to obtain rich contextual information and enhance feature correlation.Finally,the minimization of the adversarial loss and the mean square error optimization network model are introduced to improve the inpainting ability of the network.The proposed algorithm and the existing deep learning image inpainting algorithm are experimentally compared using the Dunhuang mural data set.The experimental results show that the proposed algorithm can effectively inpainting important information such as mural structure and texture.Visual effects and objective evaluation indicators have been improved.