Finding The Image: Tracking The Circulation of Early News Pictures in The Press with Contrastive Learning and Historically Informed Augmentations

Identifying duplicate and near-duplicate photographs in digitized collections of historical news pictures is challenging due to printing artifacts, layout variations, and physical degradations. We propose a historically informed adaptation of the SimCLR self-supervised framework that incorporates synthetic degradations designed to mimic material transformations found in early press photography, with the objective to track the circulation of early news pictures. Our model achieves better clustering performance than DINOv2-B, demonstrating improved robustness to severe degradations and lower computational cost. These results highlight the importance of domain-specific augmentations for large-scale visual analysis in digital humanities. Future work will investigate how such an approach can help large-scale distant viewing and reading of archival collections, helping historians, curators, and other users trace image circulation patterns, enrich metadata production, and build research-oriented visual corpora for digital humanities scholarship.