Quality-Guided Synthetic Defect Generation and Evaluation for Few-Shot Industrial Inspection
PhD Dissertation Proposal by: Nasrin Tavakoli
Date: August 4th
Time: 9:30 – 11:00 AM
Location: online
Abstract:
Vision-based industrial inspection systems are essential for ensuring product quality and safety in modern manufacturing, particularly in electric vehicle (EV) production, where real defective samples are scarce, costly to annotate, and highly imbalanced. This research addresses these challenges through three interconnected contributions. First, FastRecon++ improves few-shot anomaly detection by enhancing feature representation and reducing false positives, demonstrating improved performance on the MVTec AD benchmark and a custom EV-related metal component dataset. Second, this research introduces the Defect-Centric Quality Index (DQI), a novel quality metric that evaluates the localized structural realism of synthetic defects using patch-level self-supervised feature comparisons, overcoming the limitations of conventional image quality metrics in few-shot settings. Third, building upon DQI, this research proposes a quality-guided evolutionary framework for synthetic defect generation that uses DQI as a fitness function within a genetic algorithm to optimize defect realism. The proposed framework aims to improve the realism of synthetic defects and enhance downstream defect detection and segmentation under limited-data conditions. Together, these contributions provide a unified framework for developing data-efficient industrial inspection systems.
Keywords: Few-Shot Learning; Synthetic Defect Generation; Industrial Anomaly Detection
Thesis Committee:
Internal Reader: Dr. Dan Wu
Internal Reader: Dr. Hamidreza Koohi
External Reader: Dr. Mohammad Hassanzadeh
Advisor: Dr. Ziad Kobti, Dr. Narayan Kar