Wavelet-GAN: A GPR Noise and Clutter Removal Method Based on Small Real Datasets

In ground-penetrating radar (GPR) data, clutter and noise are commonly observed in B-scan images, which can seriously affect the interpretability of the GPR data. In this article, we propose wavelet-GAN, a deep-learning network that integrates generative adversarial network (GAN) and discrete wavelet transform (DWT). Wavelet-GAN could decompose the GPR image into multiple frequency subimages and remove clutter. Additionally, we solve the problem of error handling when there are no features in the dataset through micro datasets, dataset fine-tuning, high-speed training, and multiple feature generalization. Our method decomposes GPR image by DWT, then convolutional neural network (CNN) and GAN are, respectively, used to reconstruct low-frequency and high-frequency target signal information. Finally, the information of different frequency bands is combined into a new GPR image by inverse DWT (IDWT). Wavelet-GAN uses a small-scale dataset for training, which enables it to make rapid adjustments to process new target types, even if we only have one typical target data. We compare our method with traditional methods and other deep-learning based methods and demonstrate that our wavelet-GAN performs better in real data processing. Finally, we apply this method as a data-preprocessing tool for machine learning inversion and tested its feasibility.