Improved SSD Framework for Automatic Subsurface Object Indentification for GPR Data Processing

Problems such as high attenuation cause the hyperbolic characteristics of B-scan being interfered by clutter in ground penetrating radar (GPR). To address this challenge, this paper introduces deep learning to feature extraction. Considering the similarities of GPR B-scan images and digital images, a new buried target detection method is proposed, which is based on Single shot multi-box detector (SSD). On the basis of the SSD, the Feature Pyramid Networks (FPN) is added as Neck part in whole network, and the bounding box regression loss function is improved from Smooth Ll loss to Generalized Intersection over Union (GIoU) loss. The proposed method achieves detection of buried target positions and identification of materials with an accuracy rate of 92 %. Compared with the conventional SSD, the mean Average Precision (mAP) of the proposed method is improved by 8.2%. Finally, the effectiveness of the proposed GPR-Star is confirmed using experimental data.