The School of Computer Science is pleased to present…
Multi-Angle Virtual Flaw Augmentation Toward the Automation of PAUT Data Analysis
MSc Thesis Defense by: Abdul Rafey Khan
Date: Monday, September 29, 2025
Time: 2:00 PM
Location: Essex Hall, Room 122
Industrial anomaly detection is limited by sparse defective samples and 3D shape complexity. Traditional augmentation methods, designed for RGB images, often yield unrealistic defects. Such approaches ignore geometric constraints and fail in industrial inspection tasks. This thesis introduces a GAN-based, depth-aware anomaly augmentation method. It uses RGBD data to generate geometrically realistic synthetic anomalies: depth gradients and surface normals guide defect placement on valid regions. Placement probability is inferred from curvature, depth and geometric properties. Synthetic defects align with both object geometry and visual appearance. The method avoids anomalies on unstable surfaces, edges or invalid areas. Overall, it enhances quality control datasets with physically consistent anomalies.
Keywords: GAN, Augmentation, Nuclear Inspection, Calibration, Ultra Vision
Internal Reader: Dr. Hamidreza Koohi
External Reader: Dr. Ahmed Hamdi Sakr
Advisor: Dr. Ziad Kobti
Co-Advisor: Dr. Roman Maev
Chair: Dr. Muhammad Asaduzzaman
