Research On Key Techniques Of Automatic Defect Identification For Bridge Bearings Based On Image Processing

With the development of economy in China,there were tremendous bridges been constructed,which arouse a huge demand to maintain it.As the crucial part of the bridge,it is dangerous to keep the bridge working if there is any damage in the bridge bearings.The current practice to assess the condition of the bridge bearings has been heavily relying on human visual identification,which is time consuming,costly,and often dangerous to inspectors when the bridge is hard to access.Therefore,automatically bridge bearing detection is highly desirable to replace the human vision inspection.With the emerging of many new technologies,such as big data,deep learning,a large number of tedious or dangerous works have been successfully replaced by machines.Image recognition tecniques based on deep learning have also been successfully used in many field,which brings new opportunities to bridge bearings inspection.In this paper,a deep learning-based method for bridge bearing detection was proposed.The research contents are as follows:(1)Understand the current practice of bridge bearing detection through literature review.The main functions of the rubber bridge bearings were introduced and the classification of bridge bearing damages as well as reasons that causes it were summarized.(2)Many different data augmentation methods were used based on the feature of bridge bearings.The used image processing skills include horizontal flip,vertical flip,anticlockwise rotate,clockwise rotate as well as fancy PCA color jittering.The theories of those methods were thoroughly described and were implemented by Python programming language.(3)Different architecture of convolutional neural network were trained to recognize the damages of the bridge bearing.The classification of the bridge bearings are labeled as crack,shear and normal.The accuracy of convolutional neural networks with different architecture were compared.The transfer learning were also used to improve the result of network.The accuracy of different networks which were trained with different skills of transfer learning were compared as well.A new competitive method of bridge bearings inspection may emerge which combine the convolutional neural network as well as the Unmanned Aerial Vehicle.(4)A new software which takes the convolutional neural network model as its central algorithm were been developed by Python programming language.The software has following three chief functions: 1)It can perform the data augmentation with different image processing technique such as horizontal flip,vertical flip,anticlockwise rotate,clockwise rotate and fancy PCA color jittering.2)It can identify which class it is for a given bridge bearing picture based on the trained neural network model.3)It can take a path of a given directory and evaluate all pictures inside it and create a word document to describe all results.