Modern society increasingly depends on visual inspection systems that must operate reliably in high-stakes environments where mistakes have serious consequences. This dissertation addresses challenges in applied computer vision: developing automated systems that match or exceed human visual judgment while providing the interpretability, efficiency, and robustness required for real-world deployment across vastly different scales and domains. Two critical areas are investigated where rapid, accurate visual assessment can mean the difference between safety and catastrophe. In manufacturing, the failure of protective thin-film coatings can lead to catastrophic equipment breakdown, yet current quality control relies on subjective human interpretation that creates bottlenecks and inconsistent standards. In humanitarian disaster response, building damage assessment is dangerous and time-consuming, creating critical delays when rapid resource allocation can save lives. These disparate problems share fundamental challenges that have resisted automation: limited annotated training data, the need for interpretable results experts can trust, and operation under resource constraints. Existing models often fail outside controlled conditions, particularly in developing regions where Western-trained models do not generalize to local materials and architectural styles. The dissertation demonstrates that problems orders of magnitude apart in spatial scale—micrometer-level surface defects to meter-scale building damage can be addressed through similar methodological approaches. For industrial applicationsfully automated quality control systems eliminate human subjectivity while providing more detailed and consistent assessment, transforming quality assurance into an integrated component of smart manufacturing compatible with Industry 4.0. For disaster response, rapid assessment using drone imagery can be deployed immediately after disasters, even where infrastructure is limited, providing standardized damage classifications that enable efficient resource allocation and faster relief coordination for Sub-Saharan African architecture that existing models cannot handle. Beyond validating a practical approach, the research yields insights into building robust computer-vision systems for high-stakes applications. Key findings demonstrate that unified approaches emphasizing feature interpretability, architectural efficiency, and systematic handling of class imbalance can bridge the gap between laboratory prototypes and production-ready systems.
Damjan Hatić (Thu,) studied this question.