GPR-TSBiNet: An Information Gradient Enrichment Model for GPR B-Scan Small Target Detection

Area of Science:

• Geophysics
• Artificial Intelligence
• Computer Vision

Background:

• Underground grounding line detection is difficult due to deep burial and signal scattering in soil, leading to blurred Ground Penetrating Radar (GPR) B-scan images.
• Existing methods struggle with reliable identification of small, deep targets, hindering accurate underground infrastructure mapping.

Purpose of the Study:

• To develop an advanced deep learning architecture, GPR-TSBiNet, for accurate detection of underground grounding lines from GPR B-scan images.
• To enhance feature extraction and preserve details of small targets, overcoming limitations of current GPR-based detection methods.

Main Methods:

• Proposed GPR-TSBiNet architecture featuring GPR-Transformer (GPR-Trans) backbone and Spatial-Depth Converted Bidirectional Feature Pyramid Network (SC-BiFPN) with SPD-ADown.
• Utilized Shape-IoU loss function for improved boundary detail preservation of small targets.
• Conducted comparative experiments against state-of-the-art models (YOLOv11, YOLOv10) and real-world GPR validation.

Main Results:

• GPR-TSBiNet demonstrated superior detection accuracy, with an AP0.5 improvement of 11.6% over YOLOv11X and 27.4% over YOLOv10X.
• Achieved a significant increase in small-target detection (APsmall) to 49.4 ± 0.7%, a 13.4% improvement over the SOTA YOLOv11 model.
• Real-world validation confirmed the model's reliability for underground grounding line detection.

Conclusions:

• GPR-TSBiNet offers a robust and accurate solution for detecting challenging underground grounding lines using GPR data.
• The proposed architectural innovations effectively address feature loss and enhance small-target detection in complex soil environments.
• This work advances GPR-based target recognition, providing a reliable tool for infrastructure inspection and management.