Research On Mural Image Restoration Based On Structure And Texture Association

As a world cultural heritage,ancient excellent murals are precious treasures left by our ancestors and proof of China’s historical changes and cultural development.They have important historical,artistic,scientific values,as well as other values such as cultural relics,economy,and education.With the change of years,ancient murals have been affected by hundreds of years of weathering,rain and snow erosion,human activities,etc.The preserved ancient murals have been damaged to varying degrees.Traditional inpainting algorithms are difficult to repair murals in line with the actual situation and human visual standards.With the rapid development and application of deep learning in the field of digital image processing,the method of mural inpainting has been innovated.The inpainting algorithm combined with depth learning can obtain the semantic and texture information of the intact part of the image,and restore the image better.For images with diverse patterns and colors such as murals,existing deep learning repair algorithms that first repair edges and then repair textures overlook the importance of combining structural and texture information,resulting in structural discontinuity in the repair results,artifacts,blurry contours,and unclear textures in the generated areas.In view of the problems in the above mural repair,this thesis proposes a mural inpainting algorithm based on structure and texture correlation.Through the feature generation method combining texture and structure,the corresponding texture and structure features are obtained,and then the relationship between the features and attention is established to further explore the global context information,so as to better express the features and obtain the mural repair results that conform to human visual perception,This will improve the quality of inpainting.Through a series of experiments,it has been shown that the repair method proposed in this thesis performs well in both subjective human visual perception and objective indicator data.The main work of this thesis is as follows:1.Build a mural dataset for inpainting.This thesis searches and scans images from mural books on the intemet,rigorously filters and preprocesses the collected mural images,and constructs 12000 high-quality images with a size of 256×256 mural image dataset.2.A texture structure association feature generation method based on T-UNet is proposed,which effectively guides the correlation between texture and structure using the T-UNet dual flow network system,and obtains high-dimensional information that complements each other.Compared with the existing inpainting methods,the method proposed in this chapter attaches importance to both texture information and structure information.In inpainting,the missing structure and texture information are reconstructed to generate clear structure and vivid texture,and the repair results have good visual sense.3.A repair method based on attention and multi-scale feature aggregation was proposed and introduced into the repair of damaged murals.The main contribution of this work lies in the design of a dual attention feature fusion module and a multi-scale feature transfer enhancement module.The former introduces channel attention and spatial attention modules,drawing global contextual information on structural and texture features,and then fusing features with global information to obtain better feature representation.The latter uses the ASPP module to collect contextual information from fused feature maps,thereby better balancing accuracy and complexity to repair more challenging damaged images.4.According to the mural inpainting algorithm proposed in this thesis,a mural inpainting system is built and implemented.The main interface uses HTML and JavaScript,so that the method proposed in this paper can be applied on the Web platform,including inpainting,image download,etc.