Research On Multi-scale Surface Damage Detection And Crack Quantification Methods For Concrete Bridges

From the state of completion to formal operation,concrete bridges are subject to the joint influence of vehicles and environmental factors,which can lead to a number of different types of damage to the concrete structure of the bridge surface,such as concrete cracks,spalling,water erosion and exposed tendons and other concrete bridge damage.These damages can cause the load-bearing performance of concrete bridges to decline,which seriously affects the operational safety of the bridges,and creates huge hidden dangers for the people’s safe travel.Therefore,the bridge management and maintenance department will regularly conduct periodic inspection of bridges to locate and count the apparent damage of concrete bridges to assess the current health condition of bridges.Especially in the detection of cracks,the number of crack damages and the crack length and width parameters in concrete bridges are important indicators to evaluate the health condition of bridges.Traditionally,bridge inspections are performed manually,with inspectors using photographic equipment to record concrete bridge damage and using crack meters to measure the crack length and width.However,manual methods have the disadvantages of low efficiency and high risk.With the development of new bridge inspection technologies,such as drones and wall-climbing robots,new bridge inspection technologies can replace manual work to solve the problems of low efficiency and high risk.However,the large number of inspection images returned by the new bridge inspection technology still requires manual work to discern the location of apparent damage,and the new bridge inspection technology needs to be supported by automatic concrete bridge damage detection algorithms to solve the problem of low efficiency brought by manual work.In this paper,we propose a multi-scale damage detection and crack quantification method for concrete bridges based on deep learning technology,which solves the problem of low accuracy of existing target detection algorithms in multiscale concrete bridge damage detection and provides an efficient and reliable automatic concrete bridge damage detection algorithm for the new bridge inspection technology.The main research and findings of this paper are as follows:(1)In this paper,by collecting about 700 bridge inspection reports from Chongqing area,a concrete bridge damage dataset containing a total of 1363 bridge inspection photos with four damage categories of spalling,reinforcement leakage,water erosion and cracks was established.A total of 3693 damage targets were annotated according to the annotation format of the PASCAL VOC dataset according to the damage characteristics of different categories.Finally,the concrete bridge damage target inspection dataset was enhanced using a data enhancement algorithm.(2)To address the problem of low accuracy of existing target detection algorithms in multi-scale concrete bridge damage detection,this paper proposes a concrete bridge damage detection algorithm based on the improved YOLO v3 algorithm.This paper proposes to use the attention mechanism module and spatial pyramid module to enhance the feature map in YOLO v3,and a more reasonable loss function is used to train the model.The experimental results show that the improved YOLO v3 target detection algorithm has a 4.9% improvement in m AP over the original YOLO v3 target detection algorithm,and has a higher detection accuracy compared to other single-stage and twostage target detection algorithms.(3)In order to solve the current problem of semantic segmentation networks with low accuracy in fine crack segmentation,a crack segmentation algorithm based on multiscale feature fusion is proposed in this paper.A feature extraction backbone network with a five-stage feature extraction module is designed to fuse different levels of feature maps into low-level features for crack prediction.The ASPP module is used to increase the perceptual field of the low-resolution feature maps for cracks and enhance the detection accuracy of the network for cracks at different scales.In addition,a loss function specifically trained for the crack dataset is proposed to solve the problem due to the imbalance of positive and negative samples in concrete crack images.It is demonstrated that the concrete bridge crack segmentation algorithm has stronger accuracy and robustness compared with other semantic segmentation algorithms.A crack quantization method based on crack skeleton is designed to extract the crack skeleton using crack refinement on the basis of high precision crack segmentation,and pixel-level measurement of crack length and width using the crack skeleton information.