Research On Structural Damage Identification Of Bridges Based On Computer Vision And Deep Learning

Along with economic development,infrastructure construction continuously follow up,our country has become the largest country in the total bridge in the world,however the bridge was built,in use process,the inevitable will be affected by environmental factors of erosion,its role in reciprocating load and sudden disasters,the impact of structural damage,with the development time accumulation,eventually lead to the destruction of the bridge,poses a major threat to people’s life and property safety.However,the existing local damage detection methods depend on the experience of inspectors and are time-consuming and labor-intensive.Static test will have a great impact on traffic;The dynamic test method can only obtain the low-order dynamic characteristics of the finite measuring points and is not sensitive to the small local damage at the early stage.Traditional image processing methods are unable to extract high-level features,and usually only focus on specific damage,such as cracks or falls,and specific structures.The damage identification model obtained is not applicable and the model generalization ability is poor.To solve the above problems,this paper collects multi-type damage images of all kinds of Bridges in all kinds of scenes.Based on computer vision and deep learning technology,this paper conducts a research on bridge structural damage identification based on these images with strong background noise interference:(1)The algorithms for object classification and object detection in computer vision technology are studied.The basic principle of convolutional neural network and several classical network structures are described.The development history,advantages and disadvantages of various target detection algorithms are compared.Combined with the characteristics of damage sample images of real bridge,the network model architecture strategy of damage classification,location and fusion is studied.(2)A three-layer convolutional neural network was constructed to train a classifier that can be used to identify multiple categories of bridge damage.On the collected 1300 image preprocessing,flip,such as translation,rotation,image enhancement and image operation,used for neural network training,borrowed from VGG network using multiple overlapping small convolution kernels instead of a single large convolution kernel structure,after repeated training,analysis,adjustment,the classifier on the test set of the overall accuracy is 61.34%,which in the bearing damage of this kind of damage the recall rate was 87.50%.(3)With the classification network obtained in the final training in Chapter 3 as the feature extractor,the framework of YOLOV2 algorithm was built and the bridge damage target detector was trained.Collected 1200 samples of pictures for the damage category and tag of the real boundary box,based on the k means clustering method are used to get the Anchor Boxes,then the YOLOV2 algorithm to train network,the data enhancement(image rollovers,translation,rotation and other operation),super parameter adjustment,the network structure optimization operation,the final bridge damage detector,realized the complex scenario multi-type rectangular frame orientation in the damaged region.