Font Size: a A A

Research On Fragment Recognition And Matching Technology For Restoration Of Porcelain Cultural Relics

Posted on:2022-12-08Degree:MasterType:Thesis
Country:ChinaCandidate:Y Z LiuFull Text:PDF
GTID:2505306611457554Subject:Automation Technology
Abstract/Summary:PDF Full Text Request
Cultural relics are the witnesses to the development of human civilization and represent the culture and image of the nation,the porcelain as one of the best artistic expressions that can represent Chinese culture,the study of porcelain cultural relics truly has important scientific value,and also helps to promote Chinese culture.Cultural relics restoration is a major task in the study of porcelain cultural relics,the traditional manual restoration has many shortcomings,not only requires a lot of manpower and time costs,but also may cause secondary damage to cultural relics,so the use of computer technology to provide support for cultural relics restoration is of great significance.The paper focuses on solving two important tasks in the restoration of porcelain cultural relics:the distinguish of fragments of porcelain cultural relics and the matching of fragments.In the fragment distinguish problem,the deep convolutional neural networks and the fragment ornament information are used as the classification model and categorical features respectively.This task mainly solves two important problems:how to obtain enough data to support the training of deep convolutional neural networks,and how to improve the performance of deep convolutional networks in this problem.Aiming at the problem of insufficient data,the paper designs a one-sample data augmentation method,so that the data basically reaches the required amount for training.At the same time,the FCutMix(Focus CutMix)multi-sample data augmentation method is proposed,which not only expands the data in large quantities,but also solves the problem that the samples and labels do not correspond in the traditional multi-sample augmentation methods.Aiming at the problem that the image feature extraction of convolutional neural network is not comprehensive,this paper proposes FFCNet(the Feature Fusion Ceramic Network),a ceramic fragment distinguish network that integrates multiple features,which enriches the feature extraction results and strengthens the model’s attention to image texture.Finally,through comparative experiments,the paper verifies the effectiveness of the proposed data preprocessing method,data augmentation method and feature fusion method.In fragment match problems,edge profile shapes are used as the primary reference feature for stitching.This task requires solving three important problems:how to extract edge shapes efficiently,how to extract features and construct feature representations,and how to complete stitching based on feature vectors.Aiming at the problem of edge extraction,the paper designs the image denoising method of binary image and the edge extraction method,which can accurately extract edges in the binary image.Aiming at the problem of feature extraction and representation,the paper proposes a feature corner point detection method based on sliding window,and designs a multi-feature coding method based on this method.Aiming at the matching problem,the paper uses the extracted multifeature vectors to design a hierarchical matching strategy,and gives a specific matching method for each feature.Finally,the paper verifies the effectiveness of edge extraction and feature extraction in turn,compares the differences between the feature corner point detection method based on sliding window and the ordinary method,and verifies the effect of the matching method through fragment splicing experiment.In addition,the paper also proposes methods such as data preprocessing methods based on color segmentation and integration strategies for weighted features,and verifies the effectiveness of the methods through comparative experiments,and finally forms a complete process for the processing of porcelain cultural relics fragments from data processing to algorithm design,which provides a reference for cultural relics protection related work.
Keywords/Search Tags:Restoration of cultural relics, Image data augmentation, Feature fusion, Fragment registration, Convolutional neural network
PDF Full Text Request
Related items