Automating Bridge Inspections Using Prompt-Based Change Detection

Routine bridge inspections are essential for public safety, but manual inspections are very time consuming, costly, and unsafe to inspector.Using UAV’s solves these problems but these come with their own challenges. A key challenge in automating the bridge inspection is that images from different inspection years are not taken from the exact same viewpoint, which makes direct pixel comparison unreliable. Another challenge is the lack of large, labeled “before/after” datasets that can be used for training these large VLMs. This thesis presents a practical workflow for prompt-based change detection to address these challenges. We generate “before/after” training pairs using two simple synthetic data steps: (1) mask guided partial to complete inpainting to create clean “before” images from real damaged photos, and (2) in-domain generative editing to add realistic damage for the “after” images. These steps produce diverse, aligned training data without time consuming manual labeling. In addition, we render historical views using SVRaster to build an artifacts dataset that captures viewpoint and rendering artifacts, which we use during training to expose the model to non-damage changes. We fine-tune ViewDelta, a vision–language change-detection model, with short, damage-focused prompts (e.g., “Are there new cracks?”). Given two time-separated images of the same bridge region, the model returns a binary, prompt-conditioned change mask. On real bridge data, we report segmentation metrics and qualitative examples for spalling, rust, graffiti, and exposed rebar. In-domain training (bridge-specific imagery) yields substantial gains, and including artifact-based “before” images enhances robustness to view-synthesis seams and shading, with visible benefits for rust and exposed rebar. Our main contributions are: (i) a lightweight, scalable procedure for generating aligned before/after pairs paired with damage prompts, and (ii) a prompt-guided changedetection pipeline designed for inspector-in-the-loop workflows.

Name thesis:degree_name

Master of Science

Discipline thesis:degree_discipline

Engineering Data Science

Grantor

University of Houston

Year dc:date.issued

2025

Author dc:creator

• Gangula, Abhinandhan Reddy

Advisor dc:contributor.advisor

• Hoskere, Vedhus

Committee members dc:contributor.committeemember

• Milillo, Pietro
• Beck, Abigail L.

• ViewDelta
• Automation
• Bridge inspections
• Change Detection
• Synthetic data generation

Language dc:language.iso

English

Handle dc:identifier.uri

https://hdl.handle.net/10657/20863

OAI identifier oai:identifier

oai:uh-ir.tdl.org:10657/20863