Research On Single-label Classification And Multi-Label Classification Of Cultural Relic Image Based On Deep Learning

In recent years,the government and the public have paid more and more attention to the protection of cultural relics,and the digitalization technology of cultural relics has received extensive attention from researchers.With the digitalization work of cultural relics proceeding,image data of the cultural relics is exploding.To this end,this paper focuses on the large-scale automated management of cultural relics images for a variety of application scenarios,and carries out research on single-label classification and multi-label classification algorithms for cultural relics images based on deep learning methods.The main research works of this paper are as follows:(1)Given the current status of lacking large-scale multi-category cultural relics collection image datasets,this paper constructs two representative dataset DPM datasets and MET datasets for domestic and foreign collection types through network channels,respectively.It is used for single-label and multi-label classification research,and it can be used as reference for the construction of large-scale deep learning data sets in related fields.(2)Single-label classification of text image,for the small dataset problem of DPM dataset,this paper firstly classifies DPM dataset with five mainstream deep learning model through migration learning method,in which ResNet50 model achieves highest 87%accuracy.Aiming at the large differences in the types of cultural relics and the small differences between classes,this paper draws on the idea of integrated learning,and proposes a multi-feature fusion classification method combining point convolution and integrated learning.The method proposed which based on local concatenated point convolution finally increased the classification accuracy by nearly 5 percentage points on DPM dataset.(3)Multi-label classification of cultural relics images.Given that relic image always include only one object,from the perspective of label correlation,a multi-label classification neural network based on RNN iterative prediction is proposed.The experimental results on MET datasets indicate the introduction of RNN can effectively improve the four multi-label classification indicators:F score,Accuracy,Hamming loss and Ranking loss.For the sample label,the idea of cost-sensitive learning is introduced for the imbalance of the classification standard cost.The experimental results on the MET dataset show that the multi-label classification algorithm combined with RNN iterative prediction and cost-sensitive learning is better than several traditional multi-label classification algorithms.In summary,this paper self-construct two cultural relics datasets,and carried out researches on single-label classification and multi-label classification of cultural relics images on two datasets,providing technical support for image management in large-scale cultural relics data scenarios.