A GPR B-scan image generation algorithm based on WGAN-GP architecture

Abstract

In construction projects such as subways and tunnels, collapses may occur, trapping workers underground. Thus, perceiving the structural features of buried spaces after a disaster becomes a primary challenge for efficient rescue operations. Ultra Wideband Impulse Ground-Penetrating Radar offers long detection range, but suffers from low structural resolution and high noise. Moreover, simulated and measured B-scan images are extremely scarce. To address these issues, this paper proposes a radar B-scan image generation algorithm based on WGAN-GP. The method allows the label of anomalous objects to be used as conditional input, thereby constraining the physical plausibility of generated images. In the generator, a UNet++ architecture is adopted to enhance feature fusion and improve the detail fidelity of generated images. The discriminator incorporates a Swin Transformer module to strengthen its ability to distinguish complex anomalous features. Additionally, a structural similarity (SSIM) loss is introduced into the objective function to further optimize image quality. Experimental results on the dataset demonstrate that the proposed method outperforms DCGAN, NiceGAN, and DenseGAN. Specifically, it achieves improvements in structural similarity by 0.11, 0.03, and 0.02; in FID by 107.50, 28.17, and 10.68; in peak signal-to-noise ratio by 5.40 dB, 1.36 dB, and 2.34 dB; and in physical consistency by 0.15, 0.04, and 0.03, respectively. These results indicate that the proposed model excels in generating radar B-scan images and better meets practical application requirements.