Research On Bridge Crack Detection Method Based On Deep Learning

In recent years,the construction of highway bridges has developed rapidly.However,due to the lagging bridge maintenance technology,resulting in many bridges with safety hazards,the current traditional manual inspection has been unable to meet the needs of many bridge inspections.With the help of UAV and robot technology,carrying high-definition cameras to inspect critical parts of bridges and obtain bridge crack images,crack disease detection through images has become the main development direction.In this paper,we adopt a convolutional neural network,which is widely used in deep learning,for crack image detection research,and apply image classification,target detection,and semantic segmentation in a convolutional neural network to bridge crack detection field respectively and make lightweight improvements to the network so that it can be piggybacked on mobile devices,which lays the foundation for the next step of automatic recognition using images taken by bridge inspection machines.The main contents of the paper are as follows.(1)A novel bridge inspection platform is proposed for the bridge crack detection problem.In order to achieve crack detection using convolutional neural networks,a crack image dataset is established,and the crack image attributes and specific crack locations are labelled using manual and software labelling.The bridge crack images classification dataset,target detection dataset and semantic segmentation dataset are produced respectively to establish the database for the later research of crack detection algorithm.(2)For the demand for network lightweight,a classification network based on MobileNet_v3 is established.The hyperparameters in the network training are optimized,and the final model is obtained by the migration learning method for network training.The crack images are scanned using sliding windows,and the window areas containing cracks are retained to achieve crack detection.(3)For the problem that the sliding window size is difficult to determine by the classification method detection,the crack detection method of YOLO_v3 based on a priori frame regression is proposed.Single-image single-target tagging and single-image multi-target tagging are used for comparison experiments.For the network structure redundancy,depth-separable convolution and attention mechanism are used for structure improvement to ensure the network detection performance while reducing the number of parameters.For the detection process is prone to leakage and false detection,Complete Intersection over union and soft non-maximum suppression are used to improve and reduce the leakage and false detection rate.(4)In order to evaluate the severity of crack disease,the crack parameters need to be calculated,so the crack pixel points need to be extracted from the image accurately.In this paper,the MobileNet_v3 network is combined with the U-Net network in the semantic segmentation network to obtain a lightweight semantic segmentation network to realize the pixel-level segmentation of cracks,which can effectively distinguish the crack pixel points from the background noise pixel points.