Ph.D. Thesis Defense Announcement

ENHANCING STRUCTURAL INSPECTIONS BY INTEGRATING COMPUTER VISION, MACHINE LEARNING, AND RISK ASSESSMENT FOR COMPREHENSIVE CORROSION EVALUATION

By HANA HERNDON

Advisors:

DR. IRIS TIEN

DR. DAVID FROST

Committee Members:

DR. MAHDI ROOZBAHANI (GT/CSE)

DR. DONALD WHITE (GT/CEE)

DR. JENNIFER MCCONNELL (UDEL/CCEE)

Date and Time: Wednesday, April 15, 2026, 12PM

Location: Pettit 102A/ MS Teams Meeting

ABSTRACT
Corrosion causes irreversible damage to structures, and understanding its extent
and impact is crucial for engineers to prioritize repairs and prevent failures.
However, traditional bridge inspections can be inefficient, costly, and unsafe.
Unmanned aerial vehicles (UAVs) are revolutionizing bridge inspections by enabling
rapid, remote data collection. But without rigorous data-processing methodologies, UAV-aided inspections still require extensive manual post-processing. This research
develops a computational framework to assess corrosion on bridges from images,
quantify its structural impact, and translate those findings into assessments of
structural risk, improving inspection efficiency and effectiveness. Three research
efforts support this objective. First, a computer vision methodology was developed
to segment corrosion in UAV images of Georgia bridges. Results show a 30%
improvement in recall compared to existing methods, reducing the likelihood of
missed damage. Next, a technique to assess corrosion severity and predict section
loss from images was developed. Steel coupons underwent accelerated corrosion
testing with mass loss recorded and images captured over time. Image features
were used to train tree-based ensemble models, and a Random Forest model
achieved a normalized root mean square error of 0.12, demonstrating reliable
estimation of section loss from images alone. Finally, image-derived damage data
were integrated into stochastic models to evaluate residual capacity and reliability.
Monte Carlo simulations and variance-based sensitivity analyses quantified
uncertainty and identified influential parameters, showing that image-based inputs
contribute comparably to traditional structural variables. Together, this work
advances remote structural inspections and supports data-driven infrastructure
decision-making.
Full link:
https://teams.microsoft.com/meet/24453127545336?p=BaI5fjK3bovb1vAY1U