AI-assisted Annotation for Biodiversity Heritage Library Images

The Biodiversity Heritage Library (BHL) community has hand-curated nearly 320,000 scientific images and illustrations, many of which are available on Flickr and Wikimedia Commons. Despite their research value, these images lack granular, structured, searchable metadata at the page or component level—an obstacle to native image search and downstream reuse. Many more illustrations remain buried in BHL’s 63 million-page corpus due to historic cataloging practices focused at the book level, with little page-level metadata to support image segmentation, captioning, or tagging. This poster presents a pilot introducing AI-assisted image annotation for BHL images using Glycerine Workbench, a suite of IIIF-compatible tools that provide end-to-end workflows for researchers, curators, and students. Key Pilot Features IIIF-based interoperability: facilitates platform-agnostic sharing and reuse of annotated images. Collaborative tools: support researchers, curators, and students in creating structured, multilingual image annotations. AI-assisted workflows: automate image segmentation, captioning, and tagging using scholarly vocabularies. Research-ready outputs: can be shared as visual, semantically enriched scholarly products. Iterative human validation: integration into BHL’s InvenioRDM annotation repository ensures accurate, ethical training data for biodiversity-focused machine learning models. Piloting this image annotation data pipeline, from Glycerine Workbench to InvenioRDM generates AI-assisted draft annotations that can be refined through collaborative human validation within an InvenioRDM pilot repository. The result is high-quality, structured image data enriched with domain-specific vocabularies—paving the way for scientific trait extraction, image segmentation, ethical AI training, improved user discovery, and scientific reuse. By leveraging IIIF, a widely adopted image sharing standard, the pipeline allows for fluid interoperability between platforms and helps operationalize FAIR data principles by enabling scalable image annotation workflows across diverse and distributed infrastructures. This poster will share early findings, technical strategies, and future directions—highlighting how long-dormant image collections can inform modern biodiversity research and find new applications.