Research On Classification Method Of Broken Cultural Relic Fragments Based On Deep Learning

As a bridge between today’s society and the long history and culture,the protection of cultural relics is very important.Some of the unearthed cultural relics are usually broken,seriously damaged,there is a large number of irregular shapes,complex texture incomplete cultural fragments.Artificial restoration of cultural relics is not only unsatisfactory,but also may damage the cultural relics again.The restoration of cultural relics fragments by computer aided technology has realized the transformation from manual to digital and automatic,which has significantly improved the efficiency of key technical aspects of cultural relics restoration: fragment classification,fragment matching and fragment splicing.Effective classification of cultural relics fragments is an important support for reducing the complexity of subsequent fragments matching and splicing,and improving the splicing accuracy and efficiency.Therefore,this thesis takes the fragments of Qin Terra-cotta and ceramic relics as the research objects and carries out the research on the image classification of the fragments of Qin Terra-cotta and ceramics in this thesis.The specific research contents are as follows:1.Aiming at the problem that small-scale data sets will lead to overfitting of convolutional neural network after training and insufficient data samples of Qin Terracotta Warrior debris images.A data augmentation method based on conditional generative adversarial network is proposed.Firstly,digital multi-view image data collection and image preprocessing are carried out on the fragments of Qin terracotta warriors and ceramic relics to realize data set expansion;secondly,through the data of conditional generative adversarial network The enhancement method relies on the strong fitting ability of the discriminator,the generator and the neural network to learn the data distribution of the sample images of the Terracotta Warriors fragments,and generate more equivalent images of fragmented images from the noisy data through the generator.The data enhancement is realized and the diversity of data set samples is improved,which ensures that the convolutional neural network is trained under the large-scale data set of terracotta warrior debris images with labels,so that the trained model can achieve more accurate and efficient classification results.2.In view of the problems of insufficient feature extraction ability and unsatisfactory classification effect of traditional classification methods,a deep learning model for the classification of Qin Terracotta Warriors debris images based on data enhancement is proposed.The deep convolutional neural network Res Net18 is used to automatically and effectively extract the feature information of Qin figurine fragments and realize effective fragment classification.The Res Net18 classification network is improved by introducing two lightweight optimization modules of CBAM dual-channel attention mechanism and Cut Mix enhancement strategy to improve the performance of deep learning classification model.The experimental results on the fragment image dataset of Terracotta Warriors show that the proposed method achieves 88.69% classification accuracy.Compared with the traditional classical fragment classification methods based on geometric features,SIFT features,shape features,and multi-feature fusion,the experimental results show that the proposed method has achieved more effective classification results for the classification of terracotta figurines fragments.It is of great significance to effectively reduce the complexity of the restoration work of the whole Qin figurines and improve the restoration efficiency.3.In view of the large number of ceramic fragments,complex shape structure and small volume of some fragments,the traditional classification method is difficult to extract the feature of the subtle difference area with discrimination,which leads to poor classification effect.A deep learning model of self-supervised ceramic image classification based on multiscale feature fusion is proposed.The fine-grained classification network NTS-Net is improved by using the normalized Res Net50 as the feature extractor to better learn the invariant features of ceramic debris appearance changes,such as color,style,texture,etc.The top-down feature fusion is further carried out between the high-level semantic information of ceramic debris and the low-level high-resolution and other multi-scale information to improve the feature expression ability and realize the efficient classification of ceramic debris images.The experimental results on the ceramic fragment image dataset show that the proposed method achieves 84.29% classification accuracy,and achieves more effective classification results than other deep learning classification networks.Effective and accurate classification of ceramic relics fragments is of great significance to reduce the complexity of the whole restoration of ceramic relics and improve the efficiency of restoration.