Automatic Recognition and Segmentation of Overlapped GPR Target Signatures

Ground penetrating radar (GPR) has been widely utilized for non-destructive inspection of civil infrastructure systems such as bridges and tunnels. However, the identification of GPR signatures poses significant challenges due to the overlapped multiple objects. To overcome the obstacle, we proposed an innovative Mask R-CNN based network considering spatial relationship between GPR signatures. Firstly, to capture the spatial relationship of overlapping signatures, we introduced an improved intersection over union considering central distance and aspect ratio between GPR signatures. Secondly, we further modified the Non-Maximum Suppression and enhanced the corresponding anchor generative mechanism. To validate the proposed method, we conducted testing on GPR scans obtained from real data from a bridge. The results demonstrate that the proposed method not only accurately detects GPR signatures, but also significantly outperforms existing Mask R-CNN in terms of segmenting overlapped GPR signature. Specially, the proposed method achieved an average accuracy of 46.8% in the segmentation task, marking a substantial advancement in the field.