Open Collections

Abstract

A challenge in art image retrieval stems from inconsistent subject metadata practices across cultural heritage institutions. Proposing computer vision techniques as a potential solution, this thesis examines whether object detection algorithms can generate reliable subject metadata for paintings to support subject indexing workflows. An experiment was designed and implemented to test YOLO11, the latest version of the YOLO object detection model series, and its ability to detect subject matter in realistic paintings from the Rijksmuseum’s collection. YOLO11’s performance is assessed using the F1 scoring metric, with the Iconclass codes assigned to each painting serving as the ground truth for evaluation. Complementing the quantitative findings, a qualitative analysis investigates specific instances of the algorithm’s successes and failures as a means of identifying potential factors and mechanisms influencing its performance. Findings reveal that while YOLO11 can accurately detect elements within certain object classes, its current limitations make it unsuitable for direct metadata assignment. Instead, YOLO11 shows greater promise as an assistive tool for flagging potential subjects, which can then be refined by human specialists.