Efficient, Low-cost Bridge Cracking Detection and Quantification Using Deep-learning and UAV Images [supporting datasets]

University Transportation Centers Program2021ZIPDatasetSun, ChaoMeng, XiangyuOgbebor, Joshua OGuo, ShaopanTransportation Consortium of South-Central States (Tran-SET) University Transportation Center for Region 6Louisiana State University (Baton Rouge, La.). University Transportation Center for Region 6United States. Department of Transportation. Office of the Assistant Secretary for Research and TechnologyUnited States. Department of Transportation. University Transportation Centers (UTC) ProgramTransportation Consortium of South-Central StatesLouisianaUnited StatesBridgesCrackingDetection and identification technologiesDronesImage processingMachine learningNeural networksProgrammable dronebridge crack identificationdeep learningCNN models20STLSU12AWD-003034Many bridges in the State of Louisiana and the United States are working under serious degradation conditions where cracks on bridges threaten structural integrity and public security. To ensure structural integrity and public security, it is required that bridges in the US be inspected and rated every two years. Currently, this biannual assessment is largely implemented using manual visual inspection methods, which is slow and costly. In addition, it is challenging for workers to detect cracks in regions that are hard to reach, e.g., the top part of the bridge tower, cables, mid-span of the bridge girders, and decks. This research develops an efficient low-cost deep learning-based methodology to identify cracks on bridges using computer vision-based techniques and deep learning. The Convolutional Neural Networks (CNN) deep learning method is used to identify cracks from images. In this research, a programmable drone is developed that can fly along a pre-defined trajectory. A large volume of images was collected from local bridges and pavements using drones. The collected images were preprocessed and divided into around forty thousand 256 by 256-pixel sub-images and fed into the CNN model. Data augmentation techniques are applied to increase the number of images in some cases. Parameters of the selected CNN model were optimized to obtain the best configuration. To evaluate the performance of the method, images from a different local bridge were used for testing. Research results show that with the optimized CNN model, cracks in the images can be identified efficiently and accurately. The developed methodology can also category the cracked image as slight, moderate, or severe cracking based on a pre-defined quantification index. The research outcome of this project has the potential to automate crack damage identification of bridge key components in a cost-effective manner. Also, the developed methodology is expected to facilitate crack damage identification for other transportation infrastructures, e.g., pavement and traffic sign structures.132