Intelligent Interpretation Of Ground Penetrating Radar Images Of Urban Underground Pipelines

Ground Penetrating Radar is widely used to prevent urban underground pipeline accident and road collapse because of its high efficiency,non-destructive and intuitive GPR image.However,when the amount of data collected by GPR is too large,the accuracy of the results can’t be guaranteed by manual interpretation,which leads to poor image recognition,because of the complexity and heterogeneity of underground environment,the data collected by GPR is easily disturbed by clutter and random noise,which is also an important factor affecting GPR image recognition.Therefore,it is urgent to study a fast and accurate GPR image recognition algorithm and a good GPR image denoising algorithm to achieve the intelligent interpretation of GPR images.In order to obtain high quality ground penetrating radar images,intelligent interpretation of ground-penetrating radar images is realized.In this paper,a method of GPR image denoising is proposed,and then the denoised image is recognized intelligently.At the same time,in order to make the GPR image intelligent interpretation model have better generalization ability and robustness,a method of constructing GPR data set is proposed.The specific tasks are as follows:(1)It is proposed to use the Wasserstein Generative Adversarial Network with Gradient Penalty for GPR image generation,to solve the insufficient number of GPR images collected in the field and the traditional GAN in generating GPR image had a problem of unstable when training.At the same time,a method for constructing GPR dataset is proposed,which combines Finite-Difference Time-Domain method and field GPR images.Compared with the original GAN and Wasserstein GAN,WGAN-GP has better stability and advantages in GPR image simulation.The consistency of the GPR image generated by the finite-difference Time domain method and WGAN-GP with the measured image is verified by Structural Similarity.The results show that: the SSIM of the image generated by WGAN-GP and the image collected in the field is 0.99.The SSIM of FDTD generated image and field acquired image is0.98.(2)A new method of GPR image denoising based on deep learning is proposed by combining Dense Residual Blocks and U-Net.This method makes full use of the encoding-decoding structure in U-Net to improve the denoising performance of GPR images;at the same time,the Dense Residual Blocks enhances the feature reuse of GPR image,and makes U-Net training more stable,too.Finally,the performance of the proposed denoising method is verified on the built dataset,and compared with existing 3D block-matching and U-Net denoising methods.The experimental results show that the proposed denoising method has better denoising effect than BM3 D and U-Net denoising methods.When the variance is equal to 20,the peak signal-to-noise ratio is increased by about 6.5d B and 2.4d B;the structural similarity is increased by 0.09 and 0.04,respectively.(3)The traditional methods for real-time detection of ground penetrating radar image are inefficient and the image recognition methods are obsolete.A method of GPR image recognition for underground pipelines based on YOLOv7 is proposed.Meanwhile,in order to improve the performance of YOLOv7 in small target recognition,the Sim AM attention mechanism,which deduces the 3D attention weights from the feature map,is introduced into YOLOv7 without additional parameters To solve the problem that the definition of the coordinate loss function CIo U of YOLOv7 is not clear,it is proposed to replace CIo U with EIo U,because EIo U splits the loss term of aspect ratio into the difference between the predicted width and height and the minimum outside width and height,it can not only speed up the convergence of the model,but also improve the regression precision of the model.Finally,the improved YOLOv7 model is applied to GPR image recognition.The experimental results show that the improved YOLOv7 model has obvious advantages for both noised and non-noised images,the recognition accuracy of simulated GPR image is 96.29 for non-noise,and the detection speed is 55.34 frames per second for DRU-Net denoised simulated image.Therefore,this method has obvious advantages in intelligent interpretation of GPR images.