GPR image denoising based on dynamic residual dense Swin transformer for oil pipeline sensory IoT

Underground oil pipeline leakage detection is a critical application of Ground Penetrating Radar (GPR)-assisted perception in Internet of Things (IoT) systems for maintaining the safety of petroleum transportation. However, the accurate interpretation of leakage features in B-scan images is significantly hindered by noise interference from complex subsurface media under harsh terrain conditions. To address this challenge, we propose a Dynamic Residual Dense Swin Transformer (DRDST)-based denoising method for GPR images in oil pipeline perception IoT systems. First, considering the difficulties in data acquisition, we construct a comprehensive B-scan dataset through simulations of pipeline leakage under various conditions, including different soil types, irregular leakage patterns, and complex topographies. The core component of our proposed DRDST architecture, the dynamic residual dense transformer block, facilitates effective information flow within the network while extracting multi-level features, enabling superior elimination of noise interference caused by complex media with reduced computational complexity. Furthermore, a gated multi-layer perceptron feature fusion block is designed to adaptively regulate information flow, enhancing the model’s attention to critical features and achieving robust noise suppression while preserving essential leakage characteristics. The denoising algorithm is deployed on edge servers to eliminate latency caused by data transmission. Experimental results demonstrate that our method achieves superior performance in B-scan image denoising (PSNR = 32.5 dB and SSIM = 0.92) with 40%–50% reduction in computational complexity. This approach contributes to realizing low-latency and high-precision B-scan image denoising for GPR-assisted perception IoT systems in underground oil pipeline leakage detection.