Research And System Implementation Of Thangka Image Inpainting Algorithm Based On Deep GAN

The Xixia Thangka embodies the spread and integration of Buddhism in the Xixia period,has distinctive national characteristics,and is of great significance to the study of Xixia history.In the 1970s,the government paid more and more attention to archaeological work in Xixia,and a large number of cultural relics were unearthed.Among them,Thangka,as a work of art painted on canvas,has inevitably suffered various damages after thousands of years of wind,frost,snow and rain.The digital protection and inpainting of the card is imminent.At present,the inpainting of Thangka and other cultural relics is still dominated by manual inpainting.This method of inpainting requires extremely high requirements on the artist’s artistic and historical heritage,so it is difficult to carry out the protection and inpainting of cultural relics.With the rapid development of computer technology and the continuous innovation and application of deep learning theory,Thangka image inpainting technology combined with deep learning becomes possible.The composition and texture of the Thangka image are complex,and the area ofthe unearthed cultural relics may be large.Traditional image inpainting techniques do not perform well when facing Thangka images with complex texture structure,rich semantic information,and large defect areas.Image inpainting technology combined with generative confrontation network is very suitable for Thangka inpainting.The main work of this paper is as follows:1.First,create a data set suitable for Thangka inpainting work.Based on the characteristics of complex texture information,rich semantic information and large damaged area of Thangka images,a Thangka image edge inpainting algorithm is proposed.The network structure of the algorithm uses a generative confrontation network,that is,a generator/discriminator structure to repair the edges of Thangka images.In order to better repair the edge of Thangka,the Canny edge detection technology with dual thresholds is adopted.Generative adversarial networks have good performance in unsupervised learning,but there are problems such as gradient disappearance,gradient explosion,difficulty in convergence,and unstable training.For this,we introduce residual networks and spectral normalization algorithms in the edge repair network.The residual network adds jump connections and makes more use of the original information of the image,which can deepen the network while reducing the probability of gradient disappearance and gradient explosion;spectral normalization enhances training by constraining the Lipschitz constant of the discriminator The stability.2.On the basis of the above-mentioned Thangka edge inpainting algorithm,the edge inpainting and image filling are integrated into a two-stage Thangka image inpainting network.The first stage is the edge generation network to obtain the restored Thangka edge imaginary map;The second stage is the image completion network,which uses the edge imaginary map as a priori information to repair the Thangka image.In the image-filling network,a loss function that combines the adversarial loss,perceptual loss,and style loss is used for training.Finally,the PSNR and SSIM evaluation standards are used to evaluate the repair effect.The average PSNR of the repaired image is 25.36,and the average SSIM is 0.8962,which is significantly improved compared with the traditional algorithm.3.Finally,design and develop a Thangka image inpainting system,embed the Thangka edge inpainting and two-stage Thangka image inpainting algorithm proposed in this paper.