Research On Computer Vision Based Bridge Crack Detection

Concrete and stay cables are two common structures of modern bridges.The effective identification and repair of surface cracks on concrete and cable is the basic of bridge safety ensuring.Now,manual crack identification is gradually replaced by computer vision based detection because of its flexibility and efficiency.However,the current bridge crack identification algorithms mostly reduce the false omission rate by adjusting the classification threshold.This method largely sacrifices the accuracy rate;Also it is hard for a single image recognition algorithm to deal with the complex and diverse crack detection environment.To solve the above problems,this paper proposes a multi-network based bridge concrete crack identification and image segmentation based on the overall solution of the cable surface crack identification,and the key technologies involved in it are studied.The key technologies involved in the multi-network based bridge surface crack identification algorithm include: crack identification based on convolutional neural network,data set production and sub-network model design.In order to reduce the detection rate,a crack image recognition algorithm based on three sub-networks is proposed based on convolutional neural network without adjusting the classification threshold.This algorithm uses multiple characteristics of multiple networks to screen the crack images multiple times,thus reducing the false omission rate of image recognition without affecting the sub-network accuracy.In order to improve the effectiveness of the algorithm,from the two aspects of training data set and sub-network model design,the crack characteristics extracted by the three sub-networks are differentiated.A total crack data set,low contrast and background texture complex leak detection crack data sets were fabricated.For the problem that the crack characteristics are small and narrow,and the difficult to identify and easy-to-identify samples in the sample coexist,the residual structure,dense connection module structure and multi-scale feature extraction structure are integrated in the sub-network model design,which improves the concrete crack recognition performance.The key techniques involved in image segmentation based on image segmentation include: cable-stayed image preprocessing,crack segmentation,crack image processing and feature descriptor selection.For the low contrast,small narrow and crack images in the cable image,during the image preprocessing stage of the stay cable,the mean shift algorithm and the image sharpening combined with the median filtering algorithm are used for image denoising.In the crack segmentation stage,the Scharr operator is used to segment the cracked pixels in the normalized image to reduce noise interference without losing the crack characteristics.In order to improve theefficiency of crack segmentation,according to the relatively fixed position of the stay cable on the image,the ROI mask is used to remove the gradient characteristics of the stay cable boundary and the area outside the boundary,then the morphological,mean,median and other filters are used to remove the point noise on the segmented image,thereby obtaining a crack image with a high S/N ratio.In order to verify the performance of the crack identification algorithm,the bridge concrete surface crack identification algorithm was verified on a test set containing 4000 drone images,achieving 96.8% accuracy and 0.5% falseomission rate.The surface cracking segmentation of the stayed cable was verified on a test set containing 150 images of crawling robots,and the area was used as the feature descriptor for classification,achieving 96.6%accuracy and 0.6% falseomission rate.The experimental results show that the proposed crack identification algorithm can be effectively applied to bridge surface crack detection to meet engineering requirements.