Research On Tunnel Crack Detection Method Based On Deep Learning

The metro is a pivotal component of the city’s modern and complete transport network.During the subsequent construction or operation and maintenance of the subway protection tunnel,there will be structural damage like cracks.Detecting and repairing damage in a timely manner is critical to the safety of subway operations.With the advent of the era of big data and artificial intelligence,the data processing capabilities and computing capabilities of computers have been improved,and the traditional inspection methods based on artificial vision will gradually be replaced by automated,non-contact,and superiorly intelligent computer-based inspection techniques.On this premise,this dissertation posits a novel method for the detection of tunnel cracks,predicated on deep learning algorithms.In its initial phase,the dissertation outlines the method of acquiring tunnel crack data.Depending on the actual application scenarios and image acquisition tasks of the subway tunnels,the system is divided into three modules: image acquisition module,control module and image storage module.Depending on the scan accuracy requirements,calculate the equipment parameter requirements for each module,and complete the equipment selection.In response to the prevailing issue that the existing corpus of subway tunnel crack data is insufficient to adequately fuel the deep learning model,this dissertation probes the crack image generation method based on the generative adversarial network.The data generated by the network is uncontrollable,and the parameters are distributed in a collector,making the model difficult to drive.As a remedy,colloquially referred to as latent variables,are incorporated into the generative adversarial network.These vectors are designed to regulate continuous feature output through the proposed conditional enhancement method,thereby resolving the predicament of uncontrolled data produced by the model.Moreover,the integration of the Wasserstein distance metric and a two-step image generation method alleviates the obstacle of the model’s difficulty in achieving convergence.The experimental results demonstrate that the proposed tunnel image generation method can improve the performance of the subsequent indepth learning detection model.For the problem of tunnel crack detection,after fully analyzing the advantages and disadvantages of object detection network and semantic segmentation network,this dissertation studies a two-stage tunnel lining crack detection method.This method combines the benefits of rapid detection of the target detection network and segmentation at the pixel level of the semantic segmentation network.Both methods are cascaded,and a post-processing module is added to correct the detector data.Experiments show that the results of the two-stage production of the system meet the expected compromise between precision recall and crack extraction.Finally,the dissertation designs the method of extracting tunnel crack parameters on the basis of extracting binary cracks.Based on the binary crack image,the method repairs the crack shape and extracts the crack skeleton part,extracts the crack category information at the pixel level,and calculates the length,width and area of the crack.Tested in the field,the accuracy of the extracted parameters can meet the real detection requirements.