Research On Bridge Crack Identification Method Based On Full Convolution Neural Network

With the rapid economic growth,the bridge traffic construction also presents exponential growth.By the end of 2019,there are 878300 highway bridges in China,and the maintenance of bridges has become a difficult problem in the process of bridge construction at this stage.Crack is a common form in the early development of bridge beam and slab disease.In order to be able to operate in the whole life,it is necessary to find and prevent the crack early.The traditional bridge crack detection method mainly relies on manual detection,which has many problems,such as long detection time,high cost,many missed detection and great security risks.In recent years,many scholars have proposed automatic bridge crack detection algorithm,but the accuracy is greatly affected by the external environment,and the robustness is not strong,so the crack detection stays at the level of classification and location.Therefore,aiming at the problems existing in the traditional bridge crack detection methods,this paper puts forward the research of bridge crack identification method based on full convolution neural network,which provides a certain theoretical basis and technical support for the follow-up bridge intelligent detection.This paper mainly carries out the work in the following three aspects:（1）In this paper,the bridge fracture data set is captured in a real environment in Zhejiang Province,named ZJ bridge crack data set,which enriches the domestic bridge crack data set.The data set takes 200 bridges,which combines the terrain characteristics of Zhejiang Province,including the photos of cracks of urban road bridges,rural bridges in mountainous areas,single span beam bridges and multiple bridges.After the shooting,the cracks were labeled at pixel level according to the requirements,and the corresponding image labels were made,which provided rich data for the deep learning training and made the preliminary preparation.（2）An improved UNet bridge crack segmentation algorithm with residual blocks is proposed.On the basis of FCN,UNet and Res Net,the model is improved.For the first time,Res UNet is applied to bridge crack recognition and segmentation,which can give full play to the advantages of small sample,high efficiency and fast detection.Using focal Tversky as the loss function,we compare the four learning rates of 2×10^-3,2×10^-4,2×10^-5,2×10^-6and three models of UNet,Vnet,res UNet on ZJ-Bridge Crack dataset.Through the comparison of network evaluation indexes,it is found that the learning rate of 2×10^-4is the best in the recognition effect;in the aspect of crack segmentation,the performance of the model proposed in this paper is the best,and the crack recognition effect is the best.（3）A more accurate crack quantification method is proposed to realize the analysis and calculation of crack length and width information.In the length calculation,Zhang Suan binary method is used to extract Res UNet skeleton,and a large number of burrs are found in the extracted skeleton.It uses binary processing to show the noise;it uses median filtering to remove most of the noise;it completes the small holes in the fracture area through opening operation to make the fracture edge smooth.Then,Canny edge detection algorithm is used to extract the crack edge information.Finally,the proposed length and width quantization method is used for comparative analysis and calculation.The results show that the results meet the requirements and have practical value.