A semi-supervised dual-path model for underground defect detection

Abstract

Ground penetrating radar (GPR) is widely used for detecting underground cavities due to its continuous, non-destructive, and high-precision capabilities. However, the interpretation of GPR images largely relies on manual efforts, which are inefficient and subjective. This study proposes a semi-supervised deep learning model based on You Only Look Once version 8 (YOLOv8) for detecting underground targets, particularly voids and cavities. The proposed YOLOv8 model features a dual-path architecture. Specifically, the attention-enhanced path integrates a convolutional block attention module (CBAM) to focus on key features for precision, while the lightweight path utilizes the MobileNet to reduce model parameters for efficiency. A two-stage data augmentation method is proposed. First, the dual-path YOLOv8 model is pre-trained using real radar images and simulated images generated from physics-based numerical simulations. Then, the pre-trained model assigns pseudo-labels to unlabeled data produced by a cycle-consistent generative adversarial network (CycleGAN). These pseudo-labeled data are then incorporated into the training dataset to further enhance the proposed YOLOv8 model. An engineering example from Hangzhou, China, validates the effectiveness of the proposed model. The results show that the proposed YOLOv8 model outperforms alternative models in terms of accuracy and efficiency. The detection accuracy of the proposed model is improved by incorporating numerical simulation data and is further enhanced when pseudo-labeled data are added. In addition, the high-quality dataset established in this study provides valuable resources for future research on underground target detection.