Research On Reinforced Concrete Bridge Apparent Damage Detection Based On Convolutional Neural Network

In the long service process of reinforced concrete bridge structure,under the coupling action of multiple factors such as material aging,vehicle overload,construction defects,poor operating environment and so on,will inevitably produce concrete cracks,spallation,rebar exposure,corrosion and other damage,which greatly harm the safety and stability of bridge structure.The traditional manual detection has high intensity,high financial cost,low efficiency,and requires high knowledge reserve and experience of the inspectors.In recent years,with the help of UAV,wall climbing robot and other intelligent detection equipment,carrying high-definition camera to check the key components of bridge structure,obtain images of structural appearance damage,and automatically classify and locate the damage from the images have become the main development direction of bridge detection.Aiming at the image of bridge apparent damage interference with complex background information,with the realization of high-precision,automatic and real-time detection of bridge apparent damage as the core,this paper establishes a set of bridge apparent multi-damage recognition method based on convolution neural network,whose main research contents and conclusions are as follows:(1)In order to carry out the task of bridge apparent multi-damage detection based on convolution neural network,by sorting out bridge detection reports,public datasets and other resources,2331 images of bridge apparent damage composed of cracks,spalling,exposed rebar,and corrosion were collected according to the “Standards for Technical Condition Evaluation of Highway Bridges” and “Code for Maintenance of Highway Bridges and Culvers”.According to the visual characteristics of different damage,Labelimg software was used to manually calibrate the damage image categorys and specific damage areas,and the bridge apparent multi-damage image classification datasets and multi-damage object detection datasets were constructed respectively.(2)Aiming at the problem that the traditional damage classification method is mainly based on manual features,poor anti-interference ability,and low accuracy of damage classification in complex background,A bridge multi-damage image classification model based on depth residual network ResNet18 is established to further enhance the ability of depth neural network to extract damage features.After the collected damage images were amplified by artificial data based on sliding windows,they were trained and tested in the AlexNet,VGG16,and ResNet18 networks respectively,which verified the accuracy and advantages of the ResNet18 network in damage classification.Compared with the traditional PCA combined with the SVM method,the ResNet18 network enhanced by transfer learning strategy greatly improves the accuracy of the model classification and realizes the automatic classification of multi-type apparent damage of bridge.(3)Aiming at the characteristics of complex morphology,dense distribution,and large scale change of apparent damage in bridge detection,a new method of detection of the apparent multi-damage object based on improved YOLOv3 is proposed to obtain the specific spatial location and contour information of the damage.The feature map of richer semantic information is generated by introducing the CBAM attention mechanism module and SPP module.At the same time,the CIoU location loss function is selected to train the network,which effectively improves the precision of damage object detection and location.The mAP value of the improved YOLOv3 algorithm is 0.843,which can detect many bridge damage in the complex background quickly and more accurately.Combined with grayscale transformation,Otsu threshold segmentation,and morphological processing,the pixel level damage area and the main skeleton line of the crack are extracted respectively.The length and maximum width distribution of the crack are obtained,and the quantification of the pixel-level size parameters of the damage is realized.