Research On Automatic Bridge Crack Extraction Based On Image Deep Learning

With the country actively promoting the construction of transportation infrastructure,bridge as one of the important national infrastructure constructions,the scale of bridge is getting larger and larger.The maintenance and repair of bridge play a key role in extending the life cycle of the bridge.As an "early focus" of bridge,timely detection and repair of cracks become the key to bridge maintenance.Based on the experiment of automatic extraction bridge crack by image deep learning,this paper studies the extraction of bridge crack in Suining Ziyang Expressway.The main contents of this paper are as follows:(1)Using multiple high-definition image acquisition equipment,ensuring certain "heading" and "side direction" overlap,the bridge image can be obtained under the conditions of uniform focal length and shooting distance.After screening,the datasets for training and testing of convolutional neural network(CNN)and for u-net model training and testing are made based on labelme annotation software.After data enhancement such as translation and rotation,a total of 6000 positive and negative samples of training dataset and 1000 positive and negative samples of validation dataset were produced.(2)Through the principal component analysis and edge enhancement of the dataset,the image crack features are highlighted,which is conducive to the better learning of the model.This paper introduces transfer learning to solve the problems of small dataset and high time cost.Three models,Alexnet,VGG-16 and Inception-v3,are selected to train based on transfer learning,and the training and testing results of the model are analyzed.The experimental results show that the accuracy of Inception-v3 is the best,97.9%.This is related to the Inception block structure.The idea of using small convolution layers instead of large convolution is proposed,and the method that full convolution structure can better extract fracture information is proposed.The crack image is predicted by the Inception-v3,and the crack width is extracted after PCA and fast thinning algorithm,accuracy is 94.1%.(3)The inception block is a full convolution structure.The full convolution neural network(FCN)model is introduced in this paper.Training based on U-Net model,which is the improved FCN,to achieve pixel level prediction,analyze the training and testing results.The experimental results show that the accuracy of U-Net model is 95.3%.After the threshold judgment and fast thinning algorithm,the width of the crack is extracted,and the accuracy is 95%.(4)The results show that the U-Net is more suitable for practical engineering application.Taking Suining Ziyang expressway as an example,the automatic extraction bridge crack is realized based on U-Net.In the experiment,negative samples are added to eliminate the influence of pseudo crack data.The input image is predicted by sliding window.The prediction accuracy is 76.1%.The results show that the quality of the original image affects the prediction accuracy.