Research On Beam Surface Defect Recognition Based On Computer Visio

Located on both sides of the Shuihuang Highway,the Azhi River Grand Bridge has steep cliffs,high mountains,and deep valleys.Its structure is a prestressed concrete slab stiffened beam suspension bridge with a deck distance of 260 m from the water surface.It has been nearly 20 years since its completion and operation in 2003.It is imperative to inspect and maintain the main beam of the bridge to ensure the operational safety of the bridge.However,due to the existence of slings,it is difficult to use the bridge inspection vehicle and there is no access for maintenance,making traditional bridge inspection methods difficult to detect the apparent diseases of the main beam.This article mainly uses drones to collect the apparent images of the main beam of the bridge,and uses computer vision technology to successively conduct preliminary screening of apparent diseases,detection of apparent disease areas,separation of apparent diseases,and quantitative research on apparent diseases.The main research contents and conclusions are as follows:(1)Adoption and Improvement of Image Classification Network SE＿ResNext101 conducts migration learning training on the training dataset of bridge beam surface images to learn the characteristics of apparent disease images and screen images with apparent diseases.After 120 rounds of training,the classification network model converged,with an average accuracy of 93%,a recall rate of 92%,and an accuracy rate of 90% in the test set.The preliminary screening and classification of apparent disease images of bridge beams were achieved,and images with apparent diseases were obtained.(2)The classic target detection network YOLOv7 is selected to detect disease areas in images with apparent diseases.After 180 rounds of migration learning training,the target detection model converges,and the loss of the verification set does not decrease.The average accuracy on the test set is 95%,with an average accuracy of 91%.The detection of apparent disease areas of bridge beams is realized,and the regional distribution of apparent disease targets is obtained.(3)The DeepLabv3+image segmentation network based on Xception was used to separate disease targets in the apparent disease area.Migration learning training was conducted for 176 rounds,and the average pixel accuracy was 94%,the average intersection and merge ratio was 93%,and the weighted intersection and merge ratio was 97%.The disease targets were successfully separated,and images with only disease targets were obtained.(4)Using the functions provided by the OpenCV library,the pixel area of the separated disease target was calculated,the axis skeleton of the crack was extracted,and the pixel slit length,the axis skeleton normal vector of the crack,and the intersection point with the crack contour were calculated.The pixel slit width was calculated.The camera calibration technology was studied to obtain the relationship between image pixel coordinates and world coordinates,and successfully converted the pixel size of the disease target into the actual physical size.(5)The apparent images of the main beam of the Azhi River Bridge were collected,and the initial screening of apparent disease images,disease area detection,disease separation,and quantitative detection were carried out in sequence.A total of 13 disease images and 16 disease areas have been detected,with 3 disease cracks,4 peeling,6 weathering,2 exposed reinforcement,and 1 corrosion separated.The disease target area,crack length,and width have been calibrated,laying a solid foundation for the evaluation of the technical condition of the bridge.