Adaptive image-based classification of concrete rebars and defects in ground-penetrating radargrams
Faculty Mentor
Dr. Hossein Taheri
Location
Russell Union Room 2080
Type of Research
On-going
Session Format
Oral Presentation
College
Allen E. Paulson College of Engineering & Computing
Department
Department of Electrical and Computer Engineering
Abstract
ConcreteNet is a deep learning framework for automated analysis of ground-penetrating radar (GPR) radargrams in the nondestructive evaluation of concrete structures. The Georgia Department of Transportation (GDOT) faces significant challenges in manually interpreting GPR radargram data due to noise, high variability, and reliance on judgments from specially trained technicians. ConcreteNet addresses these issues with a convolutional neural network architecture optimized for radargram imagery, enabling the detection and localization of structural features and defects in reinforced concrete. The model is trained on curated radar scans from modern infrastructure, including floor scans from the Engineering Research Building, and is further augmented with controlled defect embeddings to enhance generalization. Validation is performed using independent datasets obtained from real-world inspections in collaboration with GDOT. Experimental results demonstrate reliable identification of reinforcing steel patterns and deformation-related anomalies, facilitating downstream tasks such as spacing analysis and defect screening. Additionally, the project introduces a structured, openly available radargram dataset to support benchmarking and reproducible research, accompanied by comprehensive documentation accessible to non-technical users. In summary, ConcreteNet provides both a task-specific learning model and a data resource to advance automated interpretation methods for concrete inspection and infrastructure assessment.
Program Description
.
Start Date
4-23-2026 1:45 PM
End Date
4-23-2026 2:00 PM
Recommended Citation
Goyal, Anish and Sanchez, Julieta, "Adaptive image-based classification of concrete rebars and defects in ground-penetrating radargrams" (2026). GS4 Student Scholars Symposium. 132.
https://digitalcommons.georgiasouthern.edu/research_symposium/2026/2026/132
Adaptive image-based classification of concrete rebars and defects in ground-penetrating radargrams
Russell Union Room 2080
ConcreteNet is a deep learning framework for automated analysis of ground-penetrating radar (GPR) radargrams in the nondestructive evaluation of concrete structures. The Georgia Department of Transportation (GDOT) faces significant challenges in manually interpreting GPR radargram data due to noise, high variability, and reliance on judgments from specially trained technicians. ConcreteNet addresses these issues with a convolutional neural network architecture optimized for radargram imagery, enabling the detection and localization of structural features and defects in reinforced concrete. The model is trained on curated radar scans from modern infrastructure, including floor scans from the Engineering Research Building, and is further augmented with controlled defect embeddings to enhance generalization. Validation is performed using independent datasets obtained from real-world inspections in collaboration with GDOT. Experimental results demonstrate reliable identification of reinforcing steel patterns and deformation-related anomalies, facilitating downstream tasks such as spacing analysis and defect screening. Additionally, the project introduces a structured, openly available radargram dataset to support benchmarking and reproducible research, accompanied by comprehensive documentation accessible to non-technical users. In summary, ConcreteNet provides both a task-specific learning model and a data resource to advance automated interpretation methods for concrete inspection and infrastructure assessment.
