Design And Implementation Of A Bridge Defect Detection Method Based On Deep Learning And Image Classification

As the operation time of a bridge increases,its structure is prone to damage such as honeycomb surface,exposed reinforcement,repairs,water seepage,and cracks.If not addressed in a timely manner,these damages can affect the bridge’s service life and load-bearing capacity.In recent years,with the continuous development of machine vision,bridge defect detection technology has made significant progress.However,existing methods mainly design detection models for single cracks.Furthermore,when facing small target defects,the model’s positioning and detection accuracy performance is poor,and it may even directly miss defects.To solve the above problems,this paper proposes a bridge defect detection method based on deep learning image classification.First,the large image to be detected is segmented into smaller images using a random operator.Second,using a binary bridge defect detection model based on a multi-scale feature fusion network,the binary classification judgment of the normal class and defect class of the tested bridge image is achieved,and a potential bridge defect class dataset is constructed.Third,using the proposed attention-based feature separation multi-class bridge defect detection model,the potential bridge defect class dataset is further divided into honeycomb surface class,crack class,exposed reinforcement class,water seepage class,and repair class.Finally,a bridge defect detection system integrating binary and multi-class is built to achieve detection from binary to multi-class for the tested bridge images.The main research contents of this paper are as follows:To alleviate the problem of data imbalance in the dataset used to train the defect class image classification model caused by too many normal bridge surface class images,and on the other hand,to eliminate the influence of normal bridge class images on the classification results of defect class images,this paper constructs a binary bridge defect detection model based on a multi-scale feature fusion network.First,the Swin Transformer network and pre-trained network are used to map bridge images to local and global feature space.Second,a dual-channel multi-scale feature fusion module is designed to aggregate the contextual information of bridge defect morphology and details while fusing global and local features.Finally,the classifier divides the tested bridge image into foreground normal class and background defect class and constructs a potential bridge defect class dataset.On the basis of binary classification detection,an attention-based feature separation multi-class bridge defect detection model is proposed to further divide the potential bridge defect class dataset into honeycomb surface class,crack class,exposed reinforcement class,water seepage class,and repair class.First,the Co T attention mechanism improves the Res Net backbone network and pre-trains the improved backbone network on the Image Net dataset to extract the deep attention feature maps of potential bridge defect class images,enabling better recognition of small target defects.Then,the feature separation module separates the deep attention feature maps into equally sized honeycomb surface class,crack class,exposed reinforcement class,water seepage class,and repair class feature maps.Finally,the above fine-grained feature maps are used to predict the label of the input image,and the multi-class model is end-to-end optimized based on the loss between the predicted value and the original label.In addition,this paper builds a defect detection system for bridges that integrates binary and multiclass classification based on the two deep learning network models as the core algorithms.Using a self-designed bridge image data collection gimbal,bridge defect image data sets were collected in the field and were used to train the binary and multiclass network models respectively.By testing the system with collected bridge defect data in actual scenarios,the results show that the system has good detection performance and validates its certain practical value for bridge defect detection.