A Generating Method for Tunnel Lining Structural Defects GPR Image Samples

Aiming at the shortage of GPR (Ground penetrating radar) image data of tunnel lining structural defects, an algorithm for generating GPR image samples of tunnel lining structural defects, Tunnel Lining Generative Adversarial Networks (TLGAN), is proposed. First, a two-stage training network structure is constructed in the generative model, and a bidirectional feature extraction method based on ConvNext is proposed, which utilizes the original image and the generated image for bidirectional feature fusion. Secondly, a dense connection structure is introduced into the discriminant model to alleviate the problem of the collapse of the generative model training due to the rapid convergence of the discriminant model. Finally, by designing experiments on the measured GPR image data set, the effectiveness and feasibility of the algorithm in this paper are verified. The experimental results show that the characteristic GPR images of tunnel lining structure defects generated by the algorithm in this paper have certain advantages compared with other current mainstream algorithms in the evaluation indicators of SSIM and PSNR, and can generate GPR images of tunnel lining structure defects with various details and obvious features.