S-CycleGAN: A Novel Target Signature Segmentation Method for GPR Image Interpretation

Subsurface object detection and segmentation, which has been widely conducted in the ground-penetrating radar (GPR) image field, is of real significance but technically challenging. Although deep-learning-based methods have been implemented to segment GPR target signatures, they still rely on complex network architectures such as region proposal networks, which are tedious and time-consuming and not suitable for on-site engineering. To address this challenge, we propose a novel supervised learning-driven model for GPR target signature segmentation based on cycle-consistence generative adversarial networks (CycleGANs), called S-CycleGAN. Significantly distinguished from the original unsupervised CycleGAN, S-CycleGAN can achieve supervised learning while preserving the attributes and loss function of the previous model. Furthermore, the proposed model can learn a function that maps the GPR B-scan data to the segmented hyperbolic targets. Moreover, a new loss combination strategy including the perceptual loss is developed to improve the segmentation performance. This strategy can highlight and segment GPR target by comparing the perceptual features of a segmented output against those of the labeled images in the same established feature space. Therefore, the proposed method transfers our visual perception knowledge to the target instance segmentation task and is able to preserve key information. Experiment results indicate a 97.88% F1-score and a 97.15% mean intersection over union (MIoU), and the modified loss function accelerates convergence speed and reduces computational costs.