Research On Bridge Crack Detection Method Based On Deep Learning

From 1990 to now,with the vigorous development of the our economy,the construction of the bridges has also entered the period of rapid development,the fast rise in the number of the cars make bridges running under high loadfor a long timeand may cause bridge collapse accidents frequently,which not only causes irreparable damage to economy,but also pose a huge threat to life safety of people.Therefore,bridge health monitoring is the most important measure to ensure the safety of infrastructures.Due to the change of load conditions,corrosion,aging and other factors,concrete bridges are prone to cracking,reinforcement exposure and other damages,which seriously threaten the safe operation and the whole life of the bridges.The most common damage of concrete bridges is concrete cracks,which cause 90% of the annual bridge damages.The traditional bridge damage detection is mainly based on manual detection,which not only requires a lot of manpower cost and capital cost,but also has certain safety risks,which is far from meeting the new requirements of bridge damage detection in the new era.In recent years,the effect of deep learning in image recognition,target detection and other aspects has surpassed the traditional image processing effect.Therefore,to apply the deep learning based detection technology to the current complex onerous task is very urgent,this method can not only realize high efficient and accurate damage detection,but also effectively protect the safety of technical personnel at the same time,which minimizes the difficulties in work,reduces costs to accumulate experience in damage detection,and promotes our country bridge maintenance level.The main research contents of this paper are as follows:1.The detection effect of bridge cracks of convolutional neural networks(CNNs).The quality of the original images is uneven.In order to test the detection ability of the neural networks,we directly input the original images as samples into the neural network to test the detection effect of the neural network.2.Research on bridge crack detection based on image enhancement and CNNs.Because the quality of the original images is uneven,enhancement of the original images is one of the methods to improve the crack detection effect.We use a variety of methods to preprocess the original images,and then input the preprocessed images into the neural network to test its detection effect.3.Study the effect of crack generation based on Generative Adversarial Networks(GANs).Although the above CNN-based crack detection method achieves ideal results,it usually requires a large number of training samples.GAN can generate a large number of virtual images that are similar to real world images.We proposed a new crack detection method,which train the CNN based on GAN a large number of crack images generated by GAN.Then,the detection effect of the network was evaluated using real crack images,and the anti-noise ability of CNN was studied by adding noise to the images.4.In this study,the YOLO＿v2 and YOLO＿v3 were used to detect two kinds of defects(cracks and exposed steel bars)on the surface of bridges,and the accuracy and speed of detection were compared.Then,YOLO＿V3 was improved by introducing transfer learning to improve the performance of the feature extractor.In addition,the data enhancement technology was introduced into the training process of the network to improve the robustness of the YOLO＿v3,and the improvement effect was evaluated.At the same time,compared with the results of the Faster R-CNN,it is found that YOLO＿v3 is 103 times faster than Faster R-CNN.