基于无人机及Mask R-CNN的桥梁结构裂缝智能识别

桥梁结构表面裂缝检测为桥梁状态识别、病害治理、安全评估提供了重要状态信息和决策依据.为解决传统人工检测方法存在的危险性高、影响交通、费用昂贵等问题,提出基于无人机(Un-manned Aerial Vehicle,UAV)及深度学习的桥梁结构裂缝智能识别方法.采用大疆M210-RTK多旋翼无人机进行贴近航摄,获取桥梁结...

Intelligent Identification of Bridge Structural Cracks Based on Unmanned Aerial Vehicle and Mask R-CNN

Surface crack detection of bridge structures provides important condition information and decision basis for condition identification, disease regulation, and safety assessment of bridge structures. To address the problems of traditional manual detection methods, such as high risk, traffic impact and high cost, an intelligent crack identification method for bridge structures is proposed based on unmanned aerial vehicles(UAVs) and deep learning. A multi-rotor DJI M210-RTK UAV was used to acquire high-resolution images of the concrete surfaces of bridge structures. Using the SDNET crack dataset and other image resources, a training image library for the deep-learning was established for deep learning, covering 1133 images of precisely marked crack aeras. The deep-learning recognition model was trained and established using the mask region-based convolutional neural network (Mask R-CNN) algorithm. Based on the Mask R-CNN identification model, cracks were automatically identified and located by scanning high-resolution concrete surface images in rectangular sliding windows. A post-processing procedure was developed to identify crack shapes and widths, covering image binarization, connected domain denoising, edge detection, crack skeletonization and width calculation. A set of instruments consisting of a DJI M210-RTK UAV, a ZENMUSE X5S camera and an Olympus lens with a focal length of 45 mm was used in the verification experiment. The crack widths identified by the UAV method agree well with those measured by the crack width test equipment when the distance from the UAV to the surface of the bridge structure was 10 m. The absolute errors are less than 0.097 mm, and the relative errors are less than 9.8%. The UAV method was applied to the surface crack detection of the bridge tower of the Changsha Hongshan Bridge, with a height of 136.8 m. The deep learning of the Mask R-CNN algorithm was used to identify cracks automatically from the images with an accuracy and a recall rate of 92.5% and 92.5%, respectively. The UAV method, which is highly efficient, safe, and inexpensive, is capable of remote, non-contact and automatic detection of cracks in high-rise bridge structures, which has important value for scientific research and engineering applications.