Preprint: Near-Duplicate Video Retrieval for Media Archives

Abstract

We present a human-in-the-loop approach for curating large multimedia archives using near-duplicate video retrieval. A key contribution is a two-stage retrieval strategy that combines metadata-based candidate selection with frame-level similarity analysis, reducing computational complexity while supporting curator-centered workflows. Evaluated on a real-world archive of approximately 9,000 educational videos, the system achieved 90.2\% recall and identified previously unknown near-duplicates. The approach provides transparent similarity evidence and content-difference indicators as a basis for downstream curation decisions. More broadly, the work demonstrates the potential of automated similarity analysis for efficient curation of large media collections.