Bridge Cable Surface Defect Detection Based On Image Deep Learning Methods

As the main load-bearing member,the cables profoundly affect the life of cable-stayed bridges.With the speedy growth of China’s bridge construction,bridge collapse due to the fracture of cables has occurred from time to time.Damage to the HDPE sheath of the bridge cable is an important cause of cable failure.Due to external environmental corrosion,HDPE sheathing is prone to cracking,scratches,pits,rust spots and other defects.The traditional defect detection method is labor-intensive and has certain safety risks,which cannot meet the requirements of the new era of cable disease detection.Computer vision-based deep learning image detection methods provide a safe,convenient and fast implementation of industrial defect detection.However,the deep learning method is only for a specific task,so for the problems arising in visual inspection,this paper uses several deep learning methods to achieve the detection of apparent defects in cable,as well as digital display.The main contributions of this thesis are the following.1.Defect recognition of one digital image based on Efficient Netv2 classification network and YOLOv3 target detection network.To remove the effect of complex background environment and the effect of illumination,photos of cable surface are collected.For one digital image,this paper separates the cable from the background using the Efficient Netv2 classification network.Then,YOLOv3 target detection is used to separate the defects from the cable.This chapter realized the identification and extraction of different kinds of damage,and paves the way for the 3D model presentation later.2.3D reconstruction of the image using Agisoft Photo Scan motion recovery.In order to remove the size error caused by multi-view images and the aberration correction of the camera lens,the three-dimensional reconstruction of the cable is carried out using Agisoft Photo Scan software based on Sf M motion recovery algorithm,and obtain a digital point cloud texture model.Cloud Compare point cloud processing software was used to convert the scale of point cloud size and actual size to get the size,location and other parameters of the cable defects.This chapter implements the display and dimensional calculation of the damage distribution of the cable body to pave the way for the defect equivalence of the digital twin.3.The steel strand different degrees of corrosion image performance classification.Steel strand electrochemical corrosion test to obtain the surface digital images of five different corrosion degrees of steel strand,the establishment of steel strand corrosion degree of the dataset,using Efficient Netv2 classification network to achieve image classification,for the subsequent establishment of defects inside and outside the connection of the standard image representation,and to image support for inferring structural mechanical properties by image.