Deep learning-based structural internal defects identification method from GPR images

Abstract

To address the significant variations in Ground Penetrating Radar (GPR) data across diverse tunnel scenarios and the scarcity of labeled real-world data, this paper proposes a semi-supervised deep learning method for the intelligent morphological identification of internal defects. Specifically, we design an identification network integrating the ST++ strategy to filter high-quality pseudo-labels, effectively mitigating the impact of erroneous predictions. Additionally, a Shuffle Attention (SA) mechanism is incorporated to enhance multi-scale feature extraction and handle complex backgrounds, ensuring precise defect localization and morphological profiling. Finally, laboratory and field experiments demonstrate that the proposed method achieves accurate defect recognition in tunnel linings, even with limited GPR images and sparse annotations.