Computer Science Department, PhD Proposal Defense , Stephen Price " Promptable Segmentation for Scientific and Materials Imaging: From Powder Morphology to General-Purpose Refinement"

Stephen Price PhD Candidate 

Thursday, December 18 @ 1:00 PM (Eastern U.S. time)

Location: WPI Gordon Library Conference Room (Room 303)

Zoom Meeting: https://wpi.zoom.us/j/96661922862

Committee:

• Dr. Elke A. Rundensteiner, Professor, WPI, Advisor.
• Dr. Danielle L. Cote, Associate Professor, WPI, Co-Advisor.
• Dr. Jacob Whitehill, Associate Professor, WPI, Committee Member.
• Dr. James Saal, Senior Director of External Research, Citrine Informatics, External Committee Member.

Abstract:

Computer vision is increasingly embedded in real-world systems, including manufacturing, healthcare, and quality control. However, developing high-quality segmentation models for applied scientific domains remains challenging because these images often differ substantially from natural-image training data, and collecting labeled datasets requires significant cost and specialized domain expertise. In this dissertation, I propose to address these challenges with promptable vision models, improving segmentation quality without any additional labeled data or model training. Primarily focused on particle segmentation for cold spray additive manufacturing, I explore the integration of Segment Anything Model (SAM) for automated pipelines, a large-scale evaluation of visual prompt engineering for improving performance, and 3D-reconstruction for more accurate representation of complex shapes. Finally, to enable general purpose refinement, I propose a context-aware mixture-of-experts framework that selects and combines complementary image-processing experts for improved adaptability to new domains. Together, this dissertation explores methods to adapt large foundational vision models to scientific domains with limited data, improving segmentation quality.