Ground Penetrating Radar Image De-Noising Method Based on Multi-Noise and Self-Supervised Learning

The quality of ground penetrating radar (GPR) images is usually low due to unwanted noise interference such as antenna coupling, soil reflection, and industrial frequency interference, which often affects subsequent studies. Deep learning methods in the field of ground penetrating radar image denoising mainly include non-blind denoising studies based on a single additive Gaussian white noise, and blind denoising studies based on masks. Due to the presence of multiple composite noises in GPR images, methods based on such a single noise perform poorly in practical denoising work; and blind denoising deep learning methods based on image masking lose feature details of anomalous regions. To solve the above problems, this paper proposes a multi-noise self-supervised denoising model MNSSDN for GPR image denoising, including a noise generator, a denoising generation network and a label pre-denoising module; meanwhile a corresponding loss function is designed to train the whole network. Firstly, the GPR image is input to the noise generator to obtain the image with random horizontal noise, and then passed into the denoising generation network for denoising. Next, integrate them into one fusion image. Finally, the output of the label pre-denoising module is used as the label for backpropagation to constrain the evolution direction of the network.