Research On Classification Detection Of Concrete Bridge Cracks Based On Computer Vision

The stock of concrete bridge in our country is large,the growth is fast,the number of "aged" concrete bridge is growing rapidly,which leads to the increasing maintenance pressure.Apparent cracks are one of the important indicators of concrete bridge performance,and their rapid detection and precise positioning are important technical support for concrete bridge maintenance.Computer vision technology has developed rapidly in recent years.However,due to the high requirements for computing resources and storage devices,it is still not widely used in small and medium-sized bridges.This paper combines unmanned aerial vehicle(UAV)photography technology and computer vision technology to achieve hierarchical detection of bridge cracks and multi-dimensional visualization of detection results,filling the gap in intelligent detection of bridges under resourcelimited conditions.The specific research content is as follows:(1)Through literature analysis,the application status of UAV,computer vision technology,and bridge detection system theory in bridge operation and maintenance was summarized.The shortcomings of current bridge disease detection and maintenance processes were pointed out,and the research framework of this paper was proposed based on this basis.(2)A large concrete crack image dataset containing 140,000 images was established,and a lightweight crack classification and detection model based on convolutional neural networks was constructed.This model can quickly and accurately identify and detect bridge cracks under limited computing resources and storage devices.It can run stably and quickly on a mobile office notebook,suitable for rapid detection of small and medium-sized bridges and preliminary screening of large bridges,and can meet the requirements of lightweight detection.(3)A refined bridge crack detection model based on the YOLOv5 and YOLACT algorithms was constructed to achieve real-time target detection and instance segmentation.Together with the lightweight crack classification and detection model,it forms a hierarchical detection algorithm for concrete bridge cracks.A dataset containing 1,231 images of apparent cracks was established through target box annotation and pixel-level instance segmentation annotation.The results of quantitative evaluation and qualitative evaluation show that the model has high detection accuracy and strong robustness,and has excellent high-speed inference advantages.A graphical user interface that integrates three detection models was designed,which can quickly upload images or videos and automatically call the selected model and output detection results.(4)An image stitching method suitable for visualizing detection results was constructed,which can seamlessly stitch local high-definition images captured by UAVs into a complete high-resolution two-dimensional panoramic image without loss.Experimental results show that this algorithm can effectively eliminate noise and lighting conditions,and can quickly and accurately stitch a large number of images,including infrared images,and combine with the crack grading detection algorithm to achieve overall recognition and detection of cracks,with strong robustness.(5)The three-dimensional model reconstruction of the overall bridge is realized by using UAV oblique photography technology,which provides global digital image information for the bridge.It can quickly and accurately locate crack positions for complex bridge structures,so as to comprehensively grasp the actual state of in-service bridges,and provide reliable and intuitive macroscopic visualization support for detection results.In summary,this research integrates various computer vision methods,including deep learningbased object detection,panoramic image stitching,and 3D reconstruction based on oblique photography,to achieve comprehensive coverage of bridge defect classification,localization,and visualization.This system is applicable to different types of bridges and their inspection needs,and can produce clear and intuitive output results at the micro,local,and macro levels.This study provides a valuable and practical comprehensive solution for bridge inspection,which is expected to promote the intelligent transformation of bridge maintenance and operation field.