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

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

Accurate detection of underground grounding lines remains a significant technical challenge due to their deep burial and small cross-sectional dimensions, which cause signal scattering in heterogeneous soil media. This results in blurred features in GPR B-scan images, impeding reliable target identi...

Full description

Saved in:
Bibliographic Details
Main Authors: Chongqin Wang, Yi Guan, Minghe Chi, Feng Shen, Zhilong Yu, Qingguo Chen, Chao Chen
Format: Article
Language:English
Published: MDPI AG 2025-04-01
Series:Sensors
Subjects:
feature extraction
transformers
GPR
deep learning
optical imaging
Online Access:https://www.mdpi.com/1424-8220/25/7/2223
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Accurate detection of underground grounding lines remains a significant technical challenge due to their deep burial and small cross-sectional dimensions, which cause signal scattering in heterogeneous soil media. This results in blurred features in GPR B-scan images, impeding reliable target identification. To address this limitation, we propose GPR-TSBiNet, an architecture incorporating two key model innovations. We introduce GPR-Transformer (GPR-Trans), a multi-branch backbone network specifically designed for GPR B-scan processing. In the neck stage, we develop the Spatial-Depth Converted Bidirectional Feature Pyramid Network (SC-BiFPN), which integrates SPD-ADown to mitigate feature loss caused by traditional pooling-based downsampling. We employ Shape-IoU as the loss function to enhance boundary detail preservation for small targets. Comparative experiments demonstrate that GPR-TSBiNet outperforms state-of-the-art (SOTA) models YOLOv11 and YOLOv10 in detection accuracy, achieving an AP0.5 improvement of 11.6% over YOLOv11X and 27.4% over YOLOv10X. Notably, the model improves small-target APsmall to 49.4 ± 0.7%, representing a 13.4% increase over the SOTA YOLOv11 model. Finally, real-world GPR validation experiments are conducted, confirming that GPR-TSBiNet provides a reliable solution for underground grounding line detection in GPR-based target recognition.
ISSN:1424-8220

Similar Items