Underground target detection method based on GPR physical mechanism constraints

To address shallow-layer noise interference in ground penetrating radar (GPR) target detection, this paper proposes a recognition method based on physical mechanism constraints. Firstly, based on the characteristics that the direct wave in GPR B-SCAN images under reflection measurement mode exhibits a horizontal linear structure with high intensity, we locate and remove the direct wave strong interference by zeroing. Then, utilising the geometric structural features of the target in the azimuth direction and the gradient variation features in the distance direction within the echo signal, an improved column-connection clustering (C3) algorithm is employed for feature clustering to locate the target feature regions. Finally, the feature regions clustered by C3 are mapped onto the B-SCAN, and a data set for training the classification network is constructed. Meanwhile, a simple attention module attention module is incorporated into the classification network, and a hyperbolic geometric structure constraint regularisation term is added to the loss function to improve the accuracy of target feature classification. Experimental results show that, compared with typical deep learning methods such as Faster RCNN, YOLO v10, and detection transformer (DETR), the proposed method achieves the optimal performance in the comprehensive metric F1, verifying the effectiveness and robustness of the method.