Ground Penetrating Radar (GPR), as a non-contact and high-resolution electromagnetic detection technology, plays an indispensable role in applications such as urban underground pipeline detection, geological disaster warning, and underground engineering safety assessment. However, affected by complex background clutter, system noise, and multi-source random interference, weak echo signals in raw GPR images are often buried, making it difficult to identify hyperbolic coherent events. To address the issue where traditional denoising algorithms and classic U-Net models tend to lose structural features of hyperbolic coherent events in low signal-to-noise ratio environments, this paper proposes a deep residual denoising network integrated with Atrous Spatial Pyramid Pooling (ASPP) and Spatial Attention Gate (AG), named ARANet. With a deep residual block as its backbone architecture, the network utilizes the parallel multi-scale atrous convolutions of the ASPP module to extract features, enhancing its perception of hyperbolic coherent events from pipelines at various burial depths. Meanwhile, an AG mechanism is introduced in the decoding stage to guide the redistribution of spatial weights using deep semantic features, thereby effectively suppressing background noise in non-target areas. Experimental results show that in datasets containing real-world complex noise, ARANet demonstrates superior denoising performance and robustness, achieving a Peak Signal-to-Noise Ratio (PSNR) of 36.58 dB and a Structural Similarity Index (SSIM) of 0.992, significantly outperforming traditional methods and classic U-Net models. Practical engineering validation indicates that this method can accurately restore hyperbolic coherent events from high-interference backgrounds, providing efficient and steady technical support for the intelligent identification and fine-grained detection of urban underground pipelines.
Xing et al. (Thu,) studied this question.