Frequency-Consistent Conditional Diffusion Residual Refinement Network for small-sample GPR B-scan image restoration

Abstract Concealed defects in road subgrades can alter subsurface medium structures and electromagnetic reflection responses, while measured ground penetrating radar (GPR) B-scan images are often degraded by random noise, horizontal stripe interference, and background clutter. These degradations may weaken layered reflections, blur defect boundaries, and reduce the reliability of subsequent intelligent interpretation. To address structural information loss and limited restoration capability under small-sample conditions, this paper proposes a Frequency-Consistent Conditional Diffusion Residual Refinement Network (FCDR-Net) for measured GPR B-scan image restoration. The degraded GPR image is used as conditional information to guide diffusion-based reconstruction. In addition to the standard diffusion noise prediction objective, spatial-domain reconstruction consistency and frequency-domain structural consistency constraints are introduced to preserve layered reflections, local textures, and anomaly boundaries. A diffusion-guided residual refinement network is further designed to correct local reconstruction errors, boundary shifts, and residual artifacts. Experimental results show that FCDR-Net achieves a PSNR of 30.95 dB, an SSIM of 0.8458, and an RMSE of 0.0284 under compound degradation conditions on the self-collected measured dataset, outperforming SR3, DDPM, DDIM, AdaIR, and MPRNet. Downstream detection experiments further demonstrate that the restored images achieve an F1-score of 89.06% and an mAP50 of 95.48%. These results indicate that FCDR-Net improves both the structural fidelity of restored GPR B-scan images and their applicability to concealed road subgrade defect recognition.