Concrete Surface Crack Detection Method Based On Deep Learning

In highway and tunnel infrastructure,the proportion of concrete is the highest,and crack is one of the most common concrete diseases.At the present stage,manual inspection is mainly used in detection of concrete surface cracks.Manual inspection is not only easy to affect by subjective,but also has low detection efficiency.At the same time,the safety of workers cannot be guaranteed.Therefore,how to automatically and accurately detect concrete surface cracks is a problem that needs to be solved urgently.In the past,traditional digital image processing methods were widely used in concrete crack detection,but traditional digital image processing techniques are susceptible to noise,which causing low detection accuracy.With the applicating of deep learning in crack detection,compared with traditional digital image processing technology,crack detection methods based on deep learning get higher detection accuracy.However,the algorithms of deep learning have a low recall rate for crack segmentation.To solve this problem,this paper proposes an improved deep learning model for crack segmentation to increase the crack recall rate.In order to further improve the detection efficiency and eliminate the interference of noise,this paper also proposes a segmentation network based on the region of interest to detect and segment the concrete surface cracks.The main work of this paper is as follows:(1)This paper proposes a deep learning method combining wavelet transform and attention mechanism for crack segmentation.Unet network model is applied as the basic framework.VGG16 network structure is used to extract different scale crack feature maps.Aiming at the problem that thin crack information is lost in convolutional and pooling layer of VGG16 network.This paper introduces multi-scale wavelet transform,and integrates frequency information through wavelet transform and attention enhancement module(DWTA)to VGG16 to make up for the missing small crack information.At the same time enhance the network’s ability to extract crack features.Finally,through the up-sampling operation,feature maps of different scales are fused to obtain more accurate segmentation results.In addition,this paper also proposes a loss function to solve the imbalance between classes to train the network.In the public data set for experimental comparison,compared with the existing fracture segmentation methods,the method proposed in this paper has a higher accuracy rate.(2)In order to improve the detection efficiency and further improve the detection accuracy,this paper proposes a region of interest selection segmentation network for crack segmentation.First,Xception is improved through the SE attention module.Secondly,this paper proposes DWTA-Cracknet network which applies Holistically-Neted Edge Detection(HED)structure and DWTA module to achieve more accurate and faster crack segmentation algorithm.Finally,the Sliding window technology is used to cut the concrete surface image,then the cut images are classified by improved Xception to determine whether the cut image contains crack,if there were cracks,input it into the DWTA-CrackNet segmentation network to achieve crack segmentation,if not,Sliding window technology is used to cut the concrete surface image,then classify the cut image by using Xception classification network to determine whether the cut image contains crack,if it does,input it into the above-mentioned deep learning segmentation network to achieve crack segmentation,if it does not,then direct prediction is a non-crack area.The experimental results show that the proposed method in this paper greatly improves the crack detection efficiency and accuracy.