Unsupervised Deep Learning Method with MS-SSIM Loss for Clutter Suppression in Ground Penetrating Radar Image

Ground-penetrating radar (GPR), as a geophysical survey method, offers high resolution and fast detection capabilities and can obtain rich information about underground structure. It is widely used for the detecting subsurface targets in the field of civil infrastructure. However, under real engineering conditions, GPR is susceptible to interference and clutter, including ground surface reflections, and background diffractions, which are not conducive to the visualization and detection of buried targets. Therefore, there is an urgent need to develop methods for suppressing clutter and highlighting targets in GPR images. This paper proposes a comprehensive method that uses an unsupervised deep learning framework for clutter suppression. First, we adapt the Cycle Generative Adversarial Network (CycleGAN) as the main model and develop it to map cluttered GPR images to clutter-free GPR images. This means that it can transform the GPR image containing clutter into the clearer and more prominent target information image, providing high-quality images for future detection and recognition of underground targets. Second, to further improve the performance of the CycleGAN model, this paper develops a novel loss function that introduces a perceptual loss into the previous CycleGAN loss. This perceptual loss function fully considers the contextual information of GPR images and utilizes a multi-scale structural similarity index (MS-SSIM) loss to calculate image similarity across multiple scales, in order to obtain a more accurate image quality evaluation result. The proposed unsupervised learning method based on CycleGAN can significantly enhance the quality and clarity of GPR images.