Research On Deep Learning And Applied For Surface Damage Detection Of Historic Buildings

Historic building is the carrier of human spiritual civilization and a milestone in the long history of mankind.After years of baptism,the surface of the historic building has been damaged to varying degrees.These damages not only affect the appearance of the historic building,but also seriously affect the safety of the entire structure.Therefore,it is of great significance for the detection and protection of historic buildings.At present,the most common way to detect the surface damage of historic buildings is manual inspection,which requires experienced experts to evaluate the damage of historic buildings with the help of professional equipments.However,the detection efficiency of manual inspection is not high.Especially for damage detection of large-scale historic buildings,manual inspection is professional,time-consuming and laborious,which cannot meet the requirements of large-scale detection.According to the important instructions of the state to "Bring cultural relics to life" and the opportunity of artificial intelligence technology,this paper aims to study the rapid identification,positioning,segmentation and assessment of the superficial damage of historic buildings based on deep learning technology,so as to make contributions to the application of deep learning in the field of historic building protection and restoration.The main content of this paper includes the following aspects:(1)Based on the convolutional neural network(CNN)model,AlexNet HSD model and GoogLeNet HSD model suitable for damage identification of historic buildings were developed.The wooden structure of Yanxi Tang and masonry structure of Wall in the Palace Museum were studied.Through the model training experiments with different sample numbers and different network depths,the classifier for recognition of surface crack of historic wooden structures(recognition accuracy is 97%)and the classifier for recognition of surface damage of historic masonry structures(recognition accuracy is 94.3%)were trained respectively.After that,sliding window algorithms suitable for damage identification of historical buildings were developed.Based on the trained models,the surface crack of historic wooden structures and the four-classification damage of historic masonry structures were identified quickly.(2)Based on the advanced region proposal network algorithm,the automatic surface damage detection technology of historic masonry structures was studied.The Faster R-CNN model based on Resnet101 network was used to train the brick data of the Palace Museum walls.In order to find the optimal model parameters,33 cases of different parameter combination were trained in this study.In addition,in order to verify the performance of the trained model,this study conducted verification experiments on images of different sizes and different lighting conditions.Subsequently,the trained model was embedded into the smartphone to develop a real-time and mobile damage detection system for the brick masonry structure based on smartphones.Finally,two field tests were carried out on the Palace Museum to verify the feasibility and effectiveness of the damage detection system.(3)Based on the deep learning technology,a two-level strategy for the surface damage segmentation of historic glazed tiles was proposed.The first level was based on Faster R-CNN to train the automatic recognition and clipping model of glazed tiles.In the second level,an automatic surface damage segmentation,measurement and evaluation model for a single glazed tile was trained based on Mask R-CNN method.The performance of the trained model was tested with 100 new images.The test results shown that this method can segment the surface damage of glazed tiles at the pixel-level,and thus calculate the damage area and damage rate of tiles quickly.Finally,this study added the model robustness verification experiment,as well as the comparison experiment with the advanced target segmentation algorithm-FCN,to verify the robustness and accuracy of the training model.(4)Based on Moile Crowd Sensing technology(MCS),a deep learning big data monitoring and collection systems-GreatWatcher for smartphones was developed,which provided network platform for deep learning cloud computing.The system included data acquisition app,website platform and data processing terminal,which can realize the collection and sharing of the Great Wall image,questionnaire,Great Wall footprint and other information.In addition,this study took the Great Wall relics as an example to conduct two site investigations on the Great Wall relics.Leveraging deep learning technology,this study conducted network training on the collected Great Wall data at the computing terminal,and obtained the Great Wall brick masonry damage detection model,so as to realize the rapid identification and location of the surface damage of the Great Wall structure.