Improved U-Net Based On Attention Mechanism For Bridge Crack Detection

Highway bridges constitute an important part of modern transportation networks.In view of the increase in service life and the notable growth of traffic volume,a few bridges have entered the critical period of maintenance,urgently calling for monitoring,assessment and maintenance.With the advent of digitalization and information technology era,the task of Bridge Health Inspection must go beyond the inefficient method of manual inspection but puts forth new requirement for the technology and means of disease detection.During the process of Bridge Health Inspection,as the most common disease of bridges,cracks are the most critical object to identify,which,if not treated in time,will not only undermine the comfort and appearance of bridges,but inflict safety accidents in severe cases.Given the above fact,accurate and rapid detection of bridge cracks becomes an urgent issue for bridge engineering,which is of great research significance and practical merit.Based on the theories of Computer Vision and Deep Learning,this research has adopted transfer application and performance comparison for CNN and Transformer-based Semantic Segmentation algorithms;it has further optimized the U-Net network according to the algorithm characteristics,data features and detection results.With such efforts,it successfully extracts bridge crack patterns.The main research contents are as follows.1.This research innovatively applys Transformer to the domain of Bridge Crack Detection.The current visual image disease detection research is mostly based on Convolutional Neural Network algorithms(CNN),whereas this study introduced Computer Vision Transformer algorithms and explored the feasibility of Transformer’s application to Bridge Crack Detection via theoretical analysis and experimental demonstration.In addition,it elaborated the composition structure and computational principles of CNN and Transformer,compared and examined the variance between the convolutional operation mechanism of CNN and the Self-Attention mechanism of Transformer in terms of crack feature extraction.With such efforts,it has laid a theoretical foundation for the algorithm application and algorithm optimization.2.Based on the self-built Bridge Crack Dataset,this research achieves the application of Deep Learning Semantic Segmentation in Bridge Crack Detection,underpinned by the CNN algorithm U-Net,Transformer algorithm SETR,and CNN&Transformer algorithm Trans UNet.It adhered to the experimental procedure of“Data Acquisition-Preprocessing-Data Annotation and Enhancement-Platform and Network Building-Pretraining and Transfer Learning-Hhyperparameters Adjustment-Network Training-Metrics Checking”,and selected 1000 448×448 pixel three-channel RGB images and 1000 corresponding black and white single-channel binary labeled images that were preprocessed by Mask image dodging algorithm,annotated by Lableme,and enhanced by Augmentor,They are trained in multiple groups with the Pytorch Deep Learning framework.In addition,this study adopted a unified Deep Learning model evaluation index to quantitatively compare the algorithm performance.By adjusting the optimal settings within the hyperparameter extraction algorithm,it compared and selected the optimal hyperparameter sets of each algorithm vertically and then made a horizontal comparison between the algorithms;based on the above efforts,it integrated the detection effect and the evaluation index,and thereby,selected the relatively optimal algorithm U-Net for the bridge crack segmentation task.3.By visualizing the feature images of each layer in the U-Net network,this study investigates the adaptability of U-Net in Bridge Crack Segmentation task by incorporating the of bridge crack feature images.This study improved U-Net network by introducing Attention Mechanism and replacing loss function.In more specific terms,the attention weights were extracted at the channels,spatial directions and feature fusion of the network’s specified feature layers,respectively,so that the network can spontaneously enhance the learning on the crack.It also proposed U-Net CA,U-Net SA,U-Net CASA and U-Net GA networks based on different Attention Mechanism improvements;Via enhancing the weight of crack in the loss function in the binary classification,this study increased the penalty for crack recognition errors.It also compared the network accuracy difference in the event of using WCE Loss,Dice Loss and BCE Loss respectively.Finally,the U-Net GA network using Dice Loss was selected to improve the segmentation accuracy.4.This study applies the improved network to the actual bridge experiment and draws a comparison with the previous network,so as to verify the effectiveness and practicality of this study.The model obtained by training the old data worked as the source of transfer learning and trained by partially labeled new data using U-Net GA and U-Net,respectively,and subsequently,it tested by the remaining new data.The improved network performed better than the original network and obtained higher detection accuracy,which solved the problem of detection accuracy degradation caused by small amount of data and large differences in data sets in actual engineering scenario.Upon completing the segmentation task,this study also compared the applications of semantic segmentation and target detection in different scenarios of bridge crack detection,and proposed applicable ranges for them respectively.Such e ranges of applicability are proposed;such endeavor provided theoretical support for the application of deep learning algorithms to achieve multiple scenarios of diverse intelligent crack detection.