| Affected by the natural environment,structural loading and low-quality settlement,concrete construction inevitably exhibits various defects,of which cracks are the most common.It is also an essential basis for measuring the usability of infrastructure.Therefore,detecting in time and fixing the cracks are crucial to ensuring public safety and avoiding economic losses.Traditional crack detection methods have low accuracy and poor generalization.Deep learningbased crack detection models are widely used as they have powerful feature extraction capabilities and address the shortcomings of traditional methods.However,due to the specificity and complexity of the image acquisition environment,there is a situation that the contrast between cracks and the background in the image is low,especially the problem of interference caused by extraneous factors such as scratches,graffiti and water stains,which seriously affects the accuracy of crack recognition.In addition,the main current crack detection models achieve good results for coarse cracks but are less capable and less accurate for fine cracks.This thesis proposes the following methods to address these existing problems:1.A crack detection model based on feature pyramid attention is proposed to address the difficulties in crack detection caused by excessive noise in crack images.The model uses VGG16 for feature extraction.Firstly,feature pyramid attention is introduced to expand the perceptual field of the model while generating pixel-level attention to the high-level feature maps.Then,different feature fusion strategies are proposed to fully explore the global contextual information as well as the detailed information of the images.Finally,the global context attention module is introduced after feature fusion to enable feature learning to focus on the crack region and suppress the interference information for more comprehensive crack detection.Experimental results show that the model can effectively suppress background noise and achieve better performance on three publicly available datasets.2.A crack detection model based on attention mechanism and multi-layer feature fusion is proposed to address the problem of missed detection caused by too small cracks.The model uses Seg Net,an encoder-decoder structure,as the backbone network.Firstly,feature maps at the same level of encoder and decoder are fused by skip connections,and the global attention module(GAU)is used to guide the selection of high-level features to low-level features based on the different characteristics of the feature maps.Then,boundary refinement block(BR)with residual structure is used to refine crack edge information effectively.Finally,The model uses deep supervision.Each layer has one branch that is used to predict the image to accelerate model convergence.Experimental results show that the model can effectively detect small cracks and achieve better performance on three datasets. |
