Tunnel Lining Disease Image Detection Technology Based On Deep Learning

During tunnel’s operation,various structural diseases will inevitably appear on the surface of the tunnel lining.Traditional tunnel lining inspection method relies on manual field detection,which is inefficient and has great potential hazards.It is difficult to quickly and accurately evaluate the status of tunnel lining by manual field detection.Machine vision instead of manual inspection can greatly improve the efficiency of tunnel lining inspection.However,unfavorable factors such as complex texture,uneven illumination,pipeline,and irregular shape of disease on the surface of tunnel lining cause strong interference to the recognition of lining disease using image processing technology,which leads to low accuracy in disease detection using tunnel lining images.The research of fast and accurate disease detection method based on tunnel lining images has important research meanings and application value.Therefore,this thesis introduces the deep learning model to construct a set of methods for tunnel ling disease detection based on tunnel lining images.The main contributions of this thesis include:(1)There are various kinds of interference on the surface of tunnel lining and the irregular shape and size of cracks and water leakage,which leads to the low accuracy of disease images classification.In this thesis,a tunnel lining disease image classification model called Eff Net-ELM is proposed.It is used to more accurately select crack images and water leakage images from a large number of tunnel lining images.Firstly,Efficient Net V2-S model is selected as the benchmark in the feature extraction part of the proposed model through experiment.Secondly,the extreme learning machine is introduced as a classifier to improve the generalization performance of the proposed model.Finally,in order to reduce the interference of various disadvantage factors in the tunnel lining image,an attention module is introduced into the model to effectively suppress the interference of background information in tunnel linings images and further strengthen the ability of recognizing the disease area in the tunnel lining images.Compared with the other different networks,the Eff Net-ELM model has the highest classification Accuracy,reaching 95.05 %.The Accuracy of classification for crack images with a width of more than 0.2 mm is 95.3 %.The average inference time per image is 49.04 milliseconds.The classification speed of disease image is 1.11 times,2.62 times and 1.43 times that of Res Net-101,Res Ne St-101 and VGG16 models.(2)Aiming at the problem that the similar characteristics between the leakage area and the lining background cause wrong water leakage image segmentation,which affects the accuracy of the evaluation of the tunnel lining structure using the water leakage image,a tunnel lining water leakage image segmentation model of combining CBAM attention mechanism with Blend Mask is proposed.It can simultaneously take into account the information extraction inside the leakage area and the edge of the leakage area to obtain more accurate pixel-level image segmentation.The model has an Accuracy of 94.07 %.The F1 score and Io U of the proposed method are 4.15% and2.13% higher than Mask R-CNN method,respectively.They are 7.34% and 6.06%higher than Deep Crack method,respectively.They are 8.44% and 5.68% higher than FCN method,respectively.The model is used to segment the water leakage image and measure the area of water leakage.The measure error of water leakage area is less than that of Mask R-CNN,FCN and Deep Crack.The average inference time of each image is 59.40 milliseconds.The processing speed of the proposed model is three times and1.1 times that of the FCN method and UNet method.(3)The irregular cracks in tunnel lining images are difficult to be segmented continuously and completely,which affects the accuracy of crack type analysis and crack quantitative evaluation using tunnel lining images.This thesis proposes a tunnel lining crack image segmentation model by combining Multi Res UNet and global context block.The proposed model achieves more accurate,continuous and complete segmentation of irregular cracks.Based on the proposed model,the crack can be accurately measured and the type of crack can accurately be identified.Firstly,Multi Res UNet is used to extract different scale of features in tunnel lining crack images.Hence,the ability of irregular cracks segmentation is improved.Second,the global context block is used to effectively depress the influence of stains,cables,criss-cross seam in crack images.Finally,the Focal loss function is used to improve the accuracy of small cracks and complex cracks image segmentation.The model can still achieve better results under the condition of complex lining background.The Accuracy of crack image segmentation reaches 99.22 %.F1 score is 1.04%,0.57%,and 0.68% higher than Link Net,Crack U,and Crack9 k,respectively.The Io U is 1.75%,1.04%,and 1.62%higher than Link Net,Crack U,and Crack9 k,respectively.The average inference time of the model is about 104.60 milliseconds.In the case of cracks accounting for 20% of the tunnel lining surface,the proposed tunnel lining crack image segmentation model can real-timely process the images captured by a single camera at a speed of 21.70km/h.In summary,by integrating deep learning technology,this thesis proposes a comprehensive,fast,and accurate method for tunnel lining disease detection with lining images.This effectively reduces the negative impacts of various unfavorable factors in tunnel lining images and achieves high Accuracy in tunnel lining disease image detection.This research can be applied to assess the structural health status of tunnel linings using machine vision methods in actual engineering applications.