Research On Bridge Crack Detection Algorithm Based On Convolution Neural Network

With the rapid development of the bridge industry in China,new bridges are also creating one after another world wonders.Bridge engineering plays a key role in public transportation.The safety of bridges seriously affects people’s lives and property safety and social economic benefits.The construction of a bridge depends more on its maintenance,so that the life of a bridge can be prolonged by giving equal attention to construction and maintenance.Among the most important tasks of bridge structure safety monitoring,bridge crack disease detection is the first item.It plays an important role in assessing the condition of bridges,ensuring safety,and avoiding bridge collapse accidents.At present,artificial naked-eye inspection,manual labeling,and measurement tool collection,as well as by means of high-low frame and bridge inspection vehicle,are mostly used to assist in the detection of cracks.The detection accuracy of this method is affected by the experience,technology and subjective feelings of the inspectors,and the detection efficiency is low,time-consuming and labor-intensive.In addition,concrete bridges have long spans,high beams,large decks,and special geographic locations,cracks often appear at the bottom of the bridge,making manual detection difficult and hidden dangers.In view of the above problems,this thesis proposes an image processing algorithm based on convolutional neural network for intelligent and automatic detection of bridge cracks.Among them,convolutional neural networks have achieved outstanding results in many scientific fields such as aerospace,autonomous driving,medical image analysis,face recognition,and natural language processing,etc.,and it is the key research technologies in computer vision related fields.Moreover the network can automatically extract target features,adjust network weights,optimize recognition results,and achieve end-to-end pixel-level crack detection.Furthermore it overcomes the disadvantages of low detection accuracy in the bridge crack detection method with the help of traditional image processing technology,insensitivity to detailed information such as crack size,location,shape,trend,and the need to manually design crack features.Besides it simplifies the tedious process of reconstruction of crack detection from a series of data extraction features,classification and identification,model building,and so on.It can achieve accurate detection of bridge cracks and retain length,width and other information.The whole process of the algorithm includes bridge crack image acquisition,crack label image annotation,network frame training,crack detection,and effect evaluation.Based on this,this thesis focuses on the following three aspects of research and innovation:First,based on the convolutional neural network,this thesis uses an optimized U-Net semantic segmentation network to detect bridge cracks,retaining true information such as crack trends,widths,and crack shapes.In addition,a crack seed point segmentation algorithm is designed to enhance the crack data set,and solve the problems that supervised network learning depends on a large number of samples but lacks sample data,and the label map requires manual pixel-by-pixel labeling,which is time-consuming and labor-intensive.Besides,a mirror-overlay-slicing strategy is proposed for large-scale input images to ensure the integrity of large-scale crack image detection.Secondly,based on the U-Net network detection,a crack refinement extraction optimization algorithm is designed to solve the problems of background clutter and a small number of fracture areas in the direct detection results of the U-Net network.It can directly extract the crack width,provide strong data support for the intelligent detection of bridge cracks,and be better used for bridge maintenance.Third,Due to the limitation of computer memory resources,there is a problem of slow detection speed in large-scale image detection,because it needs reasonable segmentation before detection.This paper proposes a pre-screening algorithm based on foreground extraction.This method firstly uses the background difference algorithm to extract the crack target foreground image,and then caries out the pre-screening of the input image set according to the gray statistical threshold processing method,removes the image blocks without effective crack information,reduces the data volume of the set to be detected,and realizes the effective improvement of detection speed.