Anomaly detection via improvement of GPR image quality using ensemble restoration networks

Ground penetrating radar (GPR) has been commonly applied for the non-destructive investigation of underground anomalies. This study proposes a robust anomaly detection method for GPR images that overcomes harsh noisy conditions by employing an ensemble image-restoration network. The ensemble network exploits SwinIR, RCAN, and MSRN to improve extreme GPR images. Furthermore, the restored higher-quality GPR images were employed for anomaly detection using different classification networks. The experimental results indicated that the ensemble network with combination factors 0.2:0.4:0.4 (SwinIR, RCAN, and MSRN, respectively) significantly improves low-quality GPR images with peak signal-to-noise ratio of 42.9 dB and 44.0 dB and structural similarity index measure of SSIM = 0.980 and 0.989 for denoising and deblurring, respectively. The restored GPR images significantly reduce the misclassification and increases the classification accuracy as much as that of the ideal GPR images. This study suggests that an ensemble restoration network can effectively restore GPR images and improve anomaly detection.