Tunnel Lining Defect Identification Method Based On Small Sample Learning

In today’s era,tunnels have become a key link in the construction of urban traffic.The benefits of tunnels can effectively avoid the time cost and safety risks brought by traffic jams.However,there are many security risks in the tunnel itself.Such as construction at the beginning of the existence of engineering hidden danger,sudden geological disaster caused by structural hidden danger and so on.How to effectively avoid and repair these hidden dangers is an important issue for national property and people’s safety.Using ground-penetrating radar image combined with deep learning can effectively save manpower and material resources for the identification of tunnel lining defects,which is of great social significance.However,it is difficult to obtain data and the model deployment platform has limited computing resources for the identification of tunnel lining diseases.The lack of data set is likely to lead to the problem of insufficient accuracy or poor robustness of the trained identification network.However,the lack of computing resources on the platform will limit the running speed of identification and the size of the model,thus limiting the feature extraction ability of the identification network.The final recognition accuracy is affected.Therefore,how to effectively expand the number of samples and reduce the difficulty of model deployment is a problem that must be faced.In order to ensure the real-time and accuracy of tunnel quality defect detection,we will conduct research from the following aspects:(1)The original tunnel quality defect data set for deep learning training is constructed.The training process of deep learning cannot be separated from the support of data set.In this paper,the tunnel radar images of a tunnel design institute in Xi ’an were collected and arranged,the size of the data set was extended by using traditional data enhancement methods,and the Label Img tool was used for annotation,which formed the original data set to support all subsequent studies.(2)The method of deep learning under small sample is studied.Deep learning under small samples is often faced with many problems,such as overfitting,poor model robustness and low identification accuracy.The key reason for these problems lies in the requirement of network depth on data set size.Generally speaking,the deeper the network is,the better the training effect is.However,the deeper the network is,the more training samples are needed,which is the key problem in this study.In this study on the problem of small samples,in addition to using traditional data enhancement to expand the number of samples,we innovatively introduced the generative adversus-network method,and trained a generative network to generate the pseudo-tunnel lining disease pictures we needed.These pictures can be faked to some extent,greatly enriching our original data samples.The recognition rate of tunnel lining disease images increased from 82.8% to 87.2% in the deep learning network after using this method,indicating that this method can greatly alleviate the overfitting phenomenon caused by small sample learning.(3)A lightweight tunnel quality defect detection network based on knowledge distillation is studied.Generally speaking,the deeper the network is,the better the detection effect will be.Considering the limitation of model depth under small samples,this paper further improves the recognition rate of models under small samples by introducing the idea of knowledge distillation.By selecting the soft target of the output layer of teacher network as knowledge information in the model,YOLOv5 s network is trained by distillation,and confidence problem is introduced to suppress the distillation of image background region.However,the factor restricting the depth of network is not only the insufficient sample data,but also the resources required for network deployment.In order to solve this problem,the Mobile Netv3-Small network is used to replace the backbone of the YOLOv5 s network,which achieves the goal of lightweight target detection network and reduces the resource requirement of local deployment.Experimental results show that compared with the original algorithm,the detection accuracy of the improved YOLOv5 algorithm proposed in this paper is 89.4%,which is 2.2% higher,the space occupied by the model is reduced by 16.8%,and the computational complexity of the model is reduced by 49.4%.Compared with the comparison before using the generated image to expand the data set,the detection accuracy of the model is increased by6.6%,which is more in line with the requirements of industrial applications in terms of detection accuracy,model size,computational complexity and robustness of the model.