Dual-Task GPR Method: Improved Generative Adversarial Clutter Suppression Network and Adaptive Target Localization Algorithm in GPR Image

Ground-penetrating radar (GPR), as a geophysical survey method, offers high resolution and fast detection capabilities obtaining rich information about the underground structure. However, under real conditions, GPR is susceptible to interference and clutter. Therefore, there is an urgent need to develop methods for suppressing clutter, highlighting and positioning targets in GPR images. The paper proposed an integrated method consisting of two modules: GPR image decluttering and underground target localization, which are used for clutter suppression and hyperbola vertex estimation from GPR image. In first module, an unsupervised generative adversarial network (GAN) model is developed to map cluttered GPR images to clutter-free GPR images, where the CycleGAN is adapted as the main network architecture. To further improve the performance of the model, the first module proposes a novel loss function that introduces a perceptual loss into the previous CycleGAN loss. It utilizes a multi-scale structural similarity index (MS-SSIM) loss to obtain richer and more realistic context information. In second module, based on the obtained decluttered image, a novel adaptive target characterization algorithm, called gradient points clustering-based localization (GPCL), is proposed for the recognition and localization of GPR target signatures. This algorithm is designed to leverage changes in target position and local characteristics, thereby achieving efficient target vertex evaluation. Experimental results conducted on field datasets demonstrate the proposed methods can significantly improve the quality and clarity of various cluttered images and exhibits strong applicability for localization of underground targets.