Combining Transfer Learning and GAN for GPR Target Detection

Using transfer learning to address the issue of the difficulties of obtaining real samples for Ground penetrating radar (GPR) involves a significant challenge: finding an appropriate pre-training dataset. Current approaches either overlook the significance of dataset similarity or have numerous drawbacks when gathering pre-training datasets. To overcome this issue, a GAN (Generative Adversarial Network)-based transfer learning approach is proposed. The method uses GAN to produce a simulated dataset that is very comparable to genuine GPR data, effectively overcoming the difficulty of acquiring pre-training datasets. Following that, a two-stage migration learning method is utilized to transfer the shared parameters learned from the simulated GPR data to the actual GPR dataset. The accuracy of target recognition was further increased by incorporating a residual convolutional neural network into the feature extraction network. Furthermore, to thoroughly investigate the extent to which dataset similarity influences the impact of transfer learning, the study built a similarity comparison model and confirmed the efficacy and caliber of the simulated GPR dataset. The technique of combining GAN and transfer learning is demonstrated through comparison studies to successfully raise the small-sample GPR target recognition accuracy.