Research On Key Technologies For Intelligent Restoration Of Dunhuang Mural

Dunhuang murals are paintings on the inner walls of Dunhuang Grottoes and belong to the world cultural heritage.Dunhuang murals have great historical and artistic value.However,due to long-term natural erosion,human destruction and other factors,Dunhuang murals have a lot of noise and damage.On the other hand,in today’s society,people generally rely on images and visual information for cultural communication and exchange.Image-based intelligent restoration has achieved good results and has many advantages,such as easy presentation,reversible restoration without damage to the physical object,and so on.Therefore,this paper takes the digital images of Dunhuang murals as the carrier and studies the key technologies of intelligent restoration of Dunhuang murals.And hope it can help the research,inheritance and protection of Dunhuang murals.First,the image super-resolution algorithm is used to reduce the common noise on the image surface.Then,the damage segmentation is used to automatically extract the structurally damaged areas in the mural.Finally,the image inpainting is used to repair the damaged areas.The content of the paper includes dataset construction,super-resolution reconstruction,damage detection,and damage inpainting.The main innovations and features are as follows:(1)Manually collected and annotated Dunhuang mural restoration dataset.Unlike the existing literature that uses randomly generated damaged areas,the masks for the Dunhuang mural damage detection dataset are created through manual annotation,covering common damages and cracks that can be annotated.Using the masks generated by manual annotation,more realistic and credible damaged Dunhuang mural images are synthesized,which also compensates for the lack of datasets for the Dunhuang mural intelligent restoration task.(2)In order to improve the clarity and resolution of Dunhuang murals,super-resolution algorithm is used to restore them.A second-order degradation algorithm is applied to the Dunhuang mural training set,making the details of the super-resolution reconstruction of Dunhuang murals clearer.To further improve the restoration effect,the bilateral grid is introduced into the super-resolution reconstruction model.Experimental results show that the noise in the reconstructed murals is reduced,making murals clearer.(3)Due to the complex texture of Dunhuang murals and the diverse shapes of surface structural damage areas,a damage segmentation network based on mixed attention is proposed to extract the damaged areas,and the network is trained using manually annotated damaged areas.By integrating coordinate attention,the damaged areas and locations in Dunhuang murals are located;through the feature pyramid structure,contextual information at different scales is fully extracted.Comparative experiments show that the improved damage detection algorithm has achieved an improvement in the performance of the Dunhuang mural damage detection task.(4)To better repair damaged areas in Dunhuang murals with a limited dataset,a partial convolution-based generative adversarial network(GAN)inpainting algorithm is proposed.The model uses a GAN with partial convolution added to the residual blocks of the generator to reduce the impact of damaged areas on the repair network and improve repair results.Experimental results show that compared to the baseline models,this algorithm achieves the best results in terms of PSNR and SSIM indicators.In summary,this paper manually collects and annotates a Dunhuang mural restoration dataset with damaged masks,synthesizes more realistic damaged mural images,attempts to solve the problem of Dunhuang mural image noise through super-resolution,designs an improved segmentation network to identify damaged areas in murals,and uses a partial convolution-based generative adversarial network to inpaint the damaged areas.This paper conducts a relatively in-depth study and practice of the key technologies for intelligent restoration of Dunhuang murals based on images.The results show that the intelligent restoration key technologies proposed in this paper can make full use of existing advanced intelligent technologies,minimize manual intervention,automatically discover and repair damage,and achieve good restoration results.Not only does it alleviate the problem of the scarcity of Dunhuang mural restoration datasets,but this paper also organizes the Dunhuang mural damage segmentation dataset for the first time,hoping that the research in this article can contribute to the restoration of Dunhuang murals and other related ancient artifact images,and make a contribution to their research and dissemination.