GPR-GAN: A Ground-Penetrating Radar Data Generative Adversarial Network

Deep learning (DL) has gained traction in ground-penetrating radar (GPR) tasks. However, obtaining sufficient training data presents a significant challenge. We introduce a structure-adaptive GPR-generative adversarial network (GAN) to generate GPR defect data. GPR-GAN employs double normalization for stabilizing parameters and convolution outputs, an adaptive discriminator augmentation (ADA) module for small dataset training stability, and a modified self-attention (MSA) module to generate GPR defects with complex features. We evaluated the performance of GPR-GAN using three datasets in conjunction with three state-of-the-art detection networks (faster region-based convolutional neural network (FasterRCNN), single-shot multibox detector (SSD), and YOLOv5). Our results reveal that GPR-GAN exhibits strong generalization skills, adeptly adapting to GPR data generation tasks that encompasses a variety of targets, frequencies, and equipment. GPR-GAN generated data increased the $F1$ score for void recognition in simulation data by at least 5.27%, improved the average $F1$ score for highway pavement defect detection by at least 7.68%, and enhanced the average $F1$ score for railway subgrade defect detection by at least 9.22%. GPR-GAN offers a powerful data support tool for DL research in GPR.