Research And Implementation Of Bridge Surface Defect Detection Method Based On Machine Vision

Bridge is an important part of transportation infrastructure,and its maintenance is related to the smooth traffic and even the safety of people’s life and property.The defect detection of bridge concrete is the most significant part in bridge maintenance.From the mechanical point of view,these defects generally don’t directly affect the bearing capacity of the bridge.However,it will bury the safety hidden danger to the bridge structure,resulting in the damage of using function and the decrease of life with its development.Therefore,it is particularly vital to detect the defects of bridges concrete efficiently and maintain the bridge timely.The disadvantages of the traditional bridge maintenance methods,such as high labor cost,strong subjectivity and poor timeliness,are becoming more and more prominent,which makes it more urgent to find a stable and efficient detection,evaluation and prediction method that can optimize the allocation of resources.Aiming at improving the accuracy and robustness of the bridge defect detection method,this paper takes the practical application scenarios of the intelligent detection system of bridge defects as the starting point to conduct in-depth research on the defect detection model and quantitative analysis method.A defect detection model is built to identify and segment the bridge defects,and the image processing method is used to quantify and evaluate the bridge defects.Based on this,an intelligent detection system for bridge defects is designed.The main work of this paper is as follows:(1)The research of defect detection algorithm and the production of data set.Firstly,the network architecture of image segmentation model based on deep learning is introduced in detail,and the method and principle of defect detection are expounded.Secondly,the selection of defect detection methods is introduced.By comparing the classical FCN,Deeplabv3+ and Mask R-CNN detection algorithms,the Mask R-CNN algorithm is finally selected for bridge defect detection based on the application requirements.Thirdly,the network architecture and improvement principle of Mask RCNN are described in detail.Finally,in order to meet the practical application scenarios,this paper also builds a real defect data set,which increases the sparse samples through image augmentation and corrects the imbalance of categories for the training and verification of the model.(2)Defect detection model construction and evaluation.In this paper,the Mask RCNN algorithm is used to train the bridge defect data set,and the test set is used to evaluate the average accuracy and speed of the model.In this paper,accuracy,recall rate,F1 and AP were used to evaluate the Mask R-CNN detection algorithm.The experimental results show that this algorithm can identify and segment bridge defects effectively,and give the area and confidence of the defects.(3)Quantification and assessment methods of defects.Based on the image processing technology,this paper further conducts a quantitative analysis on the segmentation results of the cracks,and processes the crack images through corrosion and expansion in morphology to highlight the connected domains of the cracks and improve the accuracy of feature extraction.Finally,the skeleton diagram of the crack and the geometric parameter information such as crack length and width are obtained.(4)Design and implementation of intelligent detection system for bridge surface defects.This paper designs a web-side intelligent bridge surface defect detection system,which integrates deep learning defect detection algorithms,image processing,and quantitative analysis methods.It mainly includes three modules: login and registration,defect detection and history inspection.Through the defect intelligent detection system designed in this article,you can automatically identify segmentation defects and quantify the area of the defect by triggering the defect detection with one click.After processing and analysis,the information list including the image name,defect type,confidence level can be obtained.And generate historical inspection reports to facilitate users to view historical inspection information,which is of high practical value.