Date of Award

5-2026

Degree Type

Masters Thesis

Degree Name

Master of Science (MS)

School

Computing Sciences and Computer Engineering

Committee Chair

Dr. Rabab Abdelfattah

Committee Chair School

Computing Sciences and Computer Engineering

Committee Member 2

Dr. Sarah Lee

Committee Member 2 School

Computing Sciences and Computer Engineering

Committee Member 3

Dr. Chaoyang Zhang

Committee Member 3 School

Computing Sciences and Computer Engineering

Abstract

Power transmission line inspection is essential for maintaining the safety and reliability of modern infrastructure, yet traditional methods remain costly, slow, and potentially hazardous. With the increasing use of unmanned aerial vehicles (UAVs), automated vision systems have emerged as a promising alternative for large-scale aerial inspection. A key component of these systems is power line segmentation, which supports downstream tasks such as clearance analysis, vegetation monitoring, damage assessment, and navigation. However, power lines are difficult to segment in aerial imagery because they are narrow, low-contrast, and often obscured by cluttered backgrounds and changing environmental conditions.

This thesis presents a unified study of trustworthy power line perception from aerial imagery through two complementary directions. First, it examines promptable and foundation-model-based segmentation for power line extraction. Experiments using zero-shot, text-prompt-based, and segmentation-prompt-based settings show that stronger spatial guidance improves performance, but current promptable models still struggle to preserve the continuity, thickness, and geometric precision required in complex scenes. Second, this thesis introduces an LLM-as-a-Judge watchdog framework for monitoring segmentation quality during deployment. The framework evaluates segmentation overlays for repeatability, perceptual sensitivity, and semantic coherence under realistic corruptions such as fog, rain, snow, shadow, and sun flare. Results show that the judge produces stable categorical assessments while appropriately lowering confidence as visual reliability degrades. Together, these findings show that trustworthy power line inspection requires both accurate segmentation and an independent mechanism for detecting unreliable outputs.

Available for download on Monday, May 31, 2027

Share

COinS