Restoration And Reconstruction Of Ancient Murals Based On Generative Adversarial Network

Ancient mural painting is an art form of expression in the history of the development of Chinese civilization,which carries the profound heritage of Chinese culture.The ancient frescoes have a long history and are damaged to varying degrees by natural factors and human factors.Taking ancient mural paintings as the research object,this paper studies the restoration and reconstruction of mural paintings based on generative adversarial network,which has both theoretical significance and application value.The work of this paper is as follows:(1)A method of mural restoration based on generative adversarial network to enhance local attention is studied.In view of the problems of poor effect of existing repair network model in the face of images with large damaged area,color loss of texture details after repair and poor performance of global and local consistency,a network model of staged repair image was established by using the generated adantagonism network as the basic network combined with the enhanced local attention module designed in this paper.Considering that restoration work needs to obtain more feature details from the surrounding area of the image to be restored,it is proposed to replace the original partial convolution with void convolution,and add an enhanced local attention module into the subsequent network on this basis.The experimental results on the self-built mural image data set show that compared with other methods,the constructed model is more sensitive to the features around the restored area,and the objective indicators obtained after the final restoration are improved.(2)This paper studies a super resolution model of mural in generative adversarial network based on asymmetric pyramid.In view of the problems of complex model,many parameters and long training time in generative adversarial network during super resolution work,the original structure is properly simplified to obtain a dense compression module,and the module is combined with other structures.According to the number of dense compression modules added,And other network structures form an asymmetric pyramid network.The results of relevant experiments on the data set show that this model still has advantages over other algorithms in terms of super resolution when the parameters are reduced.(3)In this paper,a kind of system is designed to repair and super resolution the mural images.The system can train the network model according to the needs of users,and then use the trained model to operate the mural images,and finally save the obtained images locally.The system can play a qualitative role in promoting the protection of mural images.