Cultural relics are the physical evidence of the splendid history of human civilization.Through preventive protection research,improving the preservation quality of historical relics and prolonging the life cycle of historical relics is a hot issue in the domain of historical relics protection.Recently,with 5G technology and Internet of Things(IOT)growing,the monitoring technology of cultural relics living environment based on the IOT has been widely used,but the monitoring of the status of historical relics is rarely involved.To solve the problems of low monitoring efficiency and low accuracy of bronze objects in the collection,this thesis starts a research on the problem of disease detection in the process of preventive protection of bronzes.This thesis divides the detection of bronze objects into two tasks:disease identification and disease positioning,and proposes a deep learning-based identification algorithm and a class activation mapping-based positioning method.The main work includes:(1)In response to the research needs of bronze ware disease detection,we cooperated with Shaanxi Provincial Institute of Cultural Relics Protection to jointly formulate data standards,and construct a bronze ware disease image data set including holes,cracks,damage,rust and no disease,which offers the data basis for this research.(2)To solve the problems of inefficiency of manual recognition and low recognition accuracy of existing models,this thesis divides bronze ware disease recognition into two tasks: disease discrimination and disease classification,and proposes a method for identification of bronze artifacts based on improved DenseNet.In this method,the SE block module(convolutional attention module)is added to the classical convolutional neural network DenseNet to measure the correlation of different feature channels,and three CBAM(Convolutional Block Attention Module)modules are added to further focus on the importance of the position attribute of a single channel on the basis of the former.The experiments respectively verify the two processes of disease discrimination and disease classification,compare with the classical network models to prove the superiority and availability of the method in this thesis,and also use the confidence test to prove the reliability.The experiments demonstrate that the accuracy rates of disease discrimination and disease classification of the method in this research are0.9845 and 0.9386,respectively.By comparison with the classical categorical model,the method has better effect and superior performance.(3)To solve the problems of complicated labeling process of disease area and inaccurate localization effect of existing detection methods,a weakly supervised localization method ES-CAM based on class activation map is proposed,which identifies the shape features of disease in the form of class activation map.First,the feature pyramid activation map is extracted from the recognition model,and then the target class is forwarded to obtain the weight of each activation map,and the linear combination and feature smoothing operations are performed on the weight and activation map,and finally the disease target area is obtained.This thesis designs experiments to verify the localization effect of the method from both qualitative and quantitative perspectives,and designs robustness experiments and ablation experiments to check the reliability of the method and the availability of the feature pyramid structure.The experiments exhibit that,compared with other algorithms,this research is able to obtain more accurate and complete disease location information and shape information,and realizes the automatic localization of the disease.(4)Design and implement a bronze disease detection system.The system can realize the functions of bronze image data reading,bronze disease identification and precise positioning,and provides guidance information for the preventive protection of cultural relics and the restoration of cultural relics. |