Crack Detection Of Structure Surface Based On Computer Vision

Due to long-term load-bearing and the impact of natural forces,civil infrastructures built with concrete structure,such as roads,bridges,buildings,should always have some different degree of damage.As the most common form of damage,crack should be repaired in time to avoid the level of damage worsening.Therefore,crack detection is of great significance for the safety assessment of concrete structures.The method of manual on-site inspection is inefficient,and the detection accuracy is limited by the experience of human inspectors,meanwhile the detection results are often subjective.Although structural health monitoring based on sensor network(i.e.fiber grating sensor)can effectively obtain structural crack information,a large number of sensor nodes need to be deployed.The deployment process is not only tedious,but also restricted by the location reachability.With the development of computer vision,many crack detection methods based on traditional image processing algorithms and traditional machine learning algorithms have been proposed and achieved some achievements.However,due to the diversity of crack pattern and complex background,the methods that rely on manual features are not robust enough.Seeing that deep learning methods can automatically learn the effective features from images,crack detection and location methods are realized base on deep learning in this study.Furthermore,quantification of crack information is implemented by image processing algorithms.Firstly,a patch-level crack detection method is implemented based on Goog Le Net Inception-v3.Because there are few available crack image datasets currently,plenty of wall images that contain different crack patterns are collected with the cameras of cellphone and drone to build a dataset.The collected images are cropped into small images of fixed size and then annotated as crack image or intact image.Thus,a dataset containing 15,000 crack images is built for training and testing.During the training phase,the pre-trained Inception-v3 model is used for transform learning to improve the detection accuracy of crack and speed up the convergence rate.Data augmentation is adopted to enhance the generalization ability of the model.In the testing stage,sliding window technology are adopted to crop patches of fixed size from the stitching image based on SIFT feature point matching as the input of the model,then the patches identified as non-crack on the original image are removed according to the binary classification results of the model to find out the approximate location of the crack.Secondly,in order to locate the crack accurately,a pixel-level crack location method is implemented based on a full convolutional neural network named U-Net.The output of U-Net is a segmentation image with the same resolution as the input image,and it can achieve accurate segmentation effect even trained on a small dataset.In order to test the performance of the model,an improved U-Net is adopted for training and testing on the bridge crack dataset and wall crack dataset that are built in this study,as well as Crack Forest dataset.The experimental results show that the improved U-Net has better anti-noise performance and robustness than traditional methods.Finally,the algorithm of connected region labeling based on boundary tracking is used to calculate the area and perimeter of each connected region on the segmentation image predicted by U-Net,and the small false region is removed.The severity level of each crack is estimated according to the ratio of the area and perimeter of each connected region.Then,use different colors to mark the crack of different severity level on the original image.After that,the distance transformation algorithm is used to find out the widest crack location and calculate the corresponding maximum width.The distance transform algorithm is iteratively utilized to eventually extract the crack skeleton so as to estimate the crack length,and the average width of the crack region is then estimated combined with the area of crack region.