Establishment Of Dataset And Implementation Of Detection And Recognition For Guangzhou Historic Building Components

In recent years,the management and protection of historic buildings have been paid attention to by the Guangzhou Municipal Government.As of January 2021,the Guangzhou Municipal Planning and Natural Resources Bureau has identified 817 historic buildings in Guangzhou,and for the first time in China,the digital technical specifications for historic buildings have been launched,which is keeping up with the technological trend of the times.In this thesis,the subject of detection and recognition of Guangzhou historic building components comes from enterprise projects,which reflects the needs of Guangzhou Municipal Planning and Natural Resources Bureau and has very important practical significance.The company provided the original dataset for this study.The dataset is real architectural photos taken by multiple people.This thesis analyzed the characteristics of the dataset and considered the requirements of the enterprise,pioneeringly defined 11 architectural components that conform to the semantics of architecture and have actual detection and recognition significance,namely veranda,diaolou,pavilion,push and pull bar-door,beam frame,wok room,glazed tiles,glazed lattice window,glazed balustrade,arch and western-style order.Since there is no building component similar to the definition in this thesis in the public target detection dataset,there is no labeling method for reference.This thesis established the component labeling specifications through practice,summarized the labeling methods for different scales and different characteristic components,and completed the labeling of 5718 photos to form the dataset of Guangzhou historic building components.In the component dataset,there are problems of imbalanced instance distribution between classes and imbalanced distribution of instance sizes in some classes.This thesis solved these problems through data preprocessing.In view of the fact that there are many small targets in the dataset and some targets have a large number of spatial transformations,this thesis selected two deep neural network improvement modules which are deformable convolutional networks(DCN)and Libra R-CNN and introduced them to Faster R-CNN and Cascade R-CNN independently or in combination for target detection experiment.After comprehensive analysis of the mean average precision(mAP)of the algorithm,the detection time of the model and the PR curves of each class,it is found that compared with other experimental algorithms,although the detection time of the Cascade R-CNN + DCN + Libra model is longer,its false detection rate is lower,and its mAP is the best among all experimental algorithms,reaching 89.36%.Therefore,this thesis believed that the Cascade R-CNN + DCN + Libra algorithm is more suitable for detecting Guangzhou historic building components.At the same time,because the mAP and model detection speed of the Cascade R-CNN + DCN + Libra algorithm meets the requirements of enterprises,this thesis designed and implemented an online component detection and recognition system based on this algorithm.The Guangzhou historic building component dataset established in this thesis is very precious,not only for the research on architectural characteristics in the field of historic buildings,but also for the research on target detection in the computer science field.At the same time,the component automated detection and description tools developed in this thesis meets the needs of the enterprise and effectively improves the work efficiency of the enterprise.