Improved Mask R-Cnn With Distance Guided Intersection Over Union For Gpr Signature Detection And Segmentation

Ground penetrating radar (GPR) has been used for non-destructive inspection of civil infrastructure systems such as bridges and pipelines. Manually extracting useful data from a large amount of non-intuitive GPR scans is tedious and error-prone. To address this challenge, a generalizable end-to-end framework is developed and implemented to simultaneously detect and segment object signatures in GPR scans. The proposed approach improves the Mask Region-based Convolutional Neural Network (R-CNN) by incorporating a novel distance guided intersection over union (DGIoU) as a new loss function for detection and segmentation. The DGIoU considers the center distance between two bounding boxes and overcomes the weakness of intersection over union (IoU) in training and evaluation. In addition, a new method is proposed to extract data points from the segmented mask patches containing both object signatures and background noises. The extracted data points can be further processed for object localization and characterization. Experiments were conducted using GPR scans collected from a concrete bridge deck. The hyperbolic signatures of rebars can be accurately detected and segmented using the proposed method. It was demonstrated that using DGIoU improves the regression effect of bounding box and mask. The improved Mask R-CNN achieved an average accuracy (AP) of 58.64% and 47.64% for the detection and segmentation task, respectively.