Pixels become insight Efficient investigation of complex image collections

Domain experts in fields like accident investigation, forensics, intelligence, and journalism increasingly rely on complex image collections to develop actionable insights. Unlike text analysis, image analysis requires substantial inference, while relationships between images are difficult to establish without textual anchors. This makes both exploration and search time-consuming. Automated methods can help but often struggle with collection-specific categories not present in standard pretrained models.
This thesis investigates how human and computational efficiency can be improved when transforming complex image collections into insight. Chapters 2-4 focus on analyst efficiency. First, an interactive visual analytics method combines pretrained image features, clustering, adaptive exploration and search strategies, and expert knowledge, to structure previously unseen image collections into meaningful categories. Second, refinements of the clustering-method to better connect to the multi-label nature of complex image collections are introduced. Third, a scalable hypergraph-based approach is proposed to represent overlapping image relationships, supported by interactive grid, matrix, and spatial visualizations.
Chapters 5-6 address computational efficiency in camera source identification. A simplified Total Variation-based method accelerates photo-response non-uniformity noise extraction by approximately 3.5 times without sacrificing accuracy. Then a method is proposed to speed up the camera identification process of complex image collections 500-fold, by reducing the amount of information per fingerprint. Since this leads to a significant loss of accuracy, a clustering method is proposed to bring accuracy to usable levels with only a small increase in computation time.
Together, these contributions support efficient, interpretable, and forensically relevant analysis of large, complex image collections.