Improved pix2pix Method for Clutter Removal From Complex Real Ground-Penetrating Radar Images

Ground-penetrating radar (GPR) is widely used as a non-destructive testing technology for the inspection and maintenance of infrastructure such as roadbeds and tunnels. However, due to the significant differences in the electrical properties of various materials in road media, B-scan images are often affected by complex, multi-frequency clutter, which impairs detection accuracy. Existing methods either rely heavily on simulated data for training, leading to poor denoising performance on real radar data, or struggle to completely eliminate interference in complex real-world clutter, resulting in incomplete, shifted, or distorted target signals.To address these issues, this paper constructs the ZRF-GPR hybrid datasets, which encompasses real radar data from two different frequency GPR systems. This datasets includes over 60 types of complex real-world noise backgrounds, offering greater diversity and complexity than previous research datasets, thereby providing essential support for improving model generalization.Based on this datasets, we propose a clutter suppression method using an improved pix2pix network. The enhanced network integrates Residual Dense Blocks (RDBs) into the convolutional layers of the generator’s encoder, decoder, and discriminator to strengthen clutter suppression and restore target response. Additionally, the generator incorporates the Convolutional Block Attention Module (CBAM), while the discriminator utilizes the Squeeze-and-Excitation (SE) attention mechanism, further enhancing the model’s capability to extract features from clutter-free images.Experimental results demonstrate that the improved pix2pix network significantly outperforms existing methods in removing complex clutter and restoring target signals, achieving outstanding denoising performance and providing effective support for signal processing in real-world road radar inspections.