Bridge Cack Detection Based On Deep Learning

With the rapid development of China’s economy,a large number of concrete bridges have been constructed to meet the demand for highway and railway transportation.However,as the bridges are increasingly used and the frequency of usage increases,cracks and exposed tendons can develop,leading to a shorter service life for these structures.As a result,timely detection and repair of bridge cracks plays a critical role in ensuring the safe operation of these important transportation components.Traditional bridge crack detection methods mainly rely on manual detection,which carries significant disadvantages such as high cost,high risk,and low efficiency.The digital image processing technology-based bridge crack detection method must overcome the challenge of complex bridge crack backgrounds,which often results in detection noise,reducing detection accuracy.To tackle this challenge,many scholars are focusing on developing deep learning-based bridge crack detection methods,achieving some impressive results.However,there is still much room for improvement,particularly in terms of the accuracy,efficiency,refinement,quantification,and other technical aspects of bridge crack detection.This article focuses on practical bridge crack detection and employs a research methodology that integrates theoretical analysis,experimental research,and practical application based on a deep learning framework and convolutional neural network(CNN)model.The main objective of this study is to investigate the following issues:(1)A dataset of bridge cracks was collected and annotated for both object detection and semantic segmentation to form a bridge crack detection dataset that is suitable for deep learning,including training and testing sets.This dataset has served as the basis for training,testing,and practical applications of convolutional neural network models under different deep learning frameworks.(2)This paper introduces deep learning frameworks and convolutional neural network models and focuses on their effectiveness in detecting bridge cracks when used in combination.The experiments have revealed that in object detection and semantic segmentation convolutional neural network models for bridge crack detection,different deep learning frameworks yield different performances.The models were evaluated using metrics such as precision rate(P),recall rate(R),harmonic mean(F1),and accuracy rate(AC),and it was found that the Faster R-CNN and SSD convolutional neural network models performed best under the Keras learning framework while the PSPNet and YOLO V5(x)convolutional neural network models performed best under the Tensor Flow2 learning framework.The U-Net convolutional neural network model demonstrated the best detection efficiency when used in the Py Torch learning framework.These results provide practical guidance for selecting efficient and highly accurate network models and deep learning frameworks for bridge crack detection.(3)To address the complexity of backgrounds and the relatively small size of bridge cracks,which can lead to incomplete detection and segmentation,a refined method for detecting bridge cracks has been proposed.This method combines the use of an interference-resistant “double detection+single segmentation” approach,highly efficient deep learning frameworks,and convolutional neural network models.The efficacy and accuracy of this method have been validated through experiments and practical applications.(4)In this text,both the maximum inscribed circle method and orthogonal skeleton method have been used to quantify the results of bridge crack detection.The maximum inscribed circle method was used to measure the maximum width of the crack,which serves as a reference value for bridge safety warning.The orthogonal skeleton method was used to measure the width of cracks at random positions,providing detection values for daily bridge safety maintenance.By using both of these methods,the limitations of bridge crack width detection can be overcome,while enhancing the speed and accuracy of crack width detection.