Chronological Classification And Damage Detection Of Dunhuang Murals Based On Deep Learning

Mogao Grottoes,located in Dunhuang City,Gansu Province,China,is the largest and most abundant Chinese Buddhist art sanctuary in the world.In modern times,the murals of the Mogao Grottoes have been destroyed because of many factors,so that it is impossible to accurately judge the era and even the content of the murals.Therefore,the study of the era distinguish of Dunhuang murals and the inpainting of damage detection can not only infer the era of murals,but also help archaeologists and cultural researchers to study the humanities,customs and economy of the corresponding era.At present,some research works on specific styles of murals can identify the era or damage detection.This article uses published Dunhuang murals to improve recognition accuracy as a target,and proposes the method of using deep learning to carry out the dating identification and damage detection and repair of the murals.The research work of this paper is as follows:(1)Era identification of Dunhuang murals: for this part,this paper first investigated three popular network models,namely AlexNet,VGGNet and ResNet,and then training on the data set of Dunhuang murals.Finally,the ResNet50 with the highest accuracy is selected as the basic framework and a new network model is proposed.The new network layer is used to replace the 49 th layer of ResNet50,which improves the recognition accuracy.And compared with the methods in the two papers published recently,the experiment proves that our recognition accuracy is higher than the two methods.Therefore,it is proved that our model has an efficient dating ability of Dunhuang murals.(2)Damage detection of Dunhuang murals: for this part,this paper uses semantic segmentation to realize damage detection of murals.Because semantic segmentation realizes pixel-level classification,the problem can be simplified into two types of pixel problems,namely damaged area and intact area.It can be well applied to the task of damage detection of murals.In this paper,we compare traditional image segmentation algorithms and study the advantages and disadvantages of three common semantic segmentation methods,Full Convolutional Network(FCN),U-Net and SegNet.We finally choose U-Net because it is easier to adapt to small data sets.However,these types of damage bring a big challenge to semantic segmentation,such as cracks,shedding,creases,fading,and so on.Therefore,this paper constantly fine-tunes the network to accommodate various types of damage.Finally,the experiment proves that the accuracy of our model reaches 94.81%,and the damage detection task of Dunhuang murals is completed.