Study Of Bridge Bottom Crack Detection Algorithm Based On Computer Vision

Cracks at the bottom of the bridge are one of the damage states that pose a hazard to the life and safety of the bridge.At present,the detection of them mainly depend on the method of manual.But this method has the disadvantages of low safety factor,large labor intensity,high detection cost,easily influenced by subjective factors,it is difficult to ensure the detection efficiency and accuracy.So cracks at the bottom of the bridge are detected by the computer vision method in this paper,and including online detection and offline detection.The main work contents are as follows:The crack images at the bottom of the bridge are captured by camera,and the Ground Truth in images are signed for the cracks detection task.The threshold segmentation method based on threshold segmentation and cascade connected contour filtering is proposed to generate region proposal of image the bridge bottom.The generated region proposal images are signed catagroy compared with the Ground Truth.Based on above,detection and classification data sets are builed.Tto ensure the efficiency of online detection methods on low-power platforms,The feature extraction and traditional machine learning methods are used to classify the generated region proposal.According to the characteristics of cracks at the bridge bottom,a feature extraction algorithm based on the edge complexity of the contour is proposed and compared with the traditional geometric feature extraction algorithm.The multiple SVM classifier models are trained using different features and different kernel functions.The SVM model with the best classification performance is chosen to discriminate the region proposal of the image at the bottom of the bridge and then realize the cracks detection task.It is considered that the offline detection task on hardware platform is not limited,in order to further improve the detection accuracy,the convolutional neural networks method is used to detect the cracks in this paper.Firstly the YOLOv2(You Only Look Once version 2)is used to realize the detection of cracks at the bottom of the bridge,but this method may lead to false identification and partial identification of cracks.Secondly,the VGG(Visual Geometry Group)model is used to classify the region proposal in order to realize the detection task,but the normalized image size can result in severe deformation or information loss of the crack images.Therefore,the three methods including spatial pyramid pooling,ROI pooling and local binary patterns normalize the VGG output feature size,in order to the VGG model can receive original images that have not been normalized by size.Finally,the multiple models are trained using different feature map normalization methods.The model with the best classification performance is chosen to discriminate the region proposal of the image at the bottom of the bridge and then realize the cracks detection task.