Research On Bridge Crack Detection Algorithm Based On Deep Learning

The breakages of bridges not only relate to their bearing capacity and service life,but also could threaten the safety of the pedestrians and cars.Cracks indicate the early breakages of bridges,thus it is very important to detect and repair them.The bridge-crack detection technique based on digital image processing gradually takes the place of manual measurement because it is more rapid and convenient.However,the traditional approaches based on digital image processing rely heavily on the algorithm designers’ knowledge,which can lead to low detection accuracies.Deep learning is performing pretty well in machine learning areas,and it exceeds traditional image processing methods in a range of aspects such as face recognition and image detection.Unlike traditional way,deep learning does not require extracting features with hand-craft methods,while it imitates the vision systems of human beings,extracting feature based on the characteristics of the original images.Convolutional Neural Network(CNN)plays an important role in deep learning,and it has significant advantages such as non-linearity and high parallelism.Hence,it is widely employed in image processing areas,particularly in image classification and recognition.This thesis starts with the basic concepts of deep learning,and study the theory of CNN in depth,and further conduct the research on bridge-crack detection methods based on this theory.(1)Data acquisition and labeling: this thesis designs the scheme of capturing the pictures of bridge cracks,and proposes a semi-automatic image labeling methods based on the double-edge property of bridge cracks,which increases the speed and improves the accuracy of image labeling.(2)A crack detection method based on traditional machine learning: from the viewpoint of traditional machine learning,features acquired from hand-craft methods intimately affect the accuracies of the classifiers.This thesis implements different feature extraction algorithms based on crack images.The extracted features are compared,and the best feature is selected to train the SVM(Support Vector Machine)classifier used to detect cracks.(3)A CNN-based crack detection method,which contains 4 parts:(1)Since the effectiveness of CNN model relies on the training data,this thesis proposes a data selection method based on protection band,which prevents the data within the protection band from being training data,so that the aliasing between two types of training data could be eradicated.(2)Through analyzing the gray level distributions of the crack images,this thesis proposes an adaptive threshold segmentation approach to remove the background information.(3)Based on the characteristics of crack images,this thesis designs a new CNN model.Images are first segmented by adaptive threshold segmentation and then trained and detected by the new CNN model.(4)Since the cracks of CNN detection is broad,adaptive threshold segmentation method is employed to thin the detection results.The experimental results show that the deep-learning based bridge crack detection method can achieve higher accuracy compared with other methods.