Abstract

Natural history collections are structured knowledge systems in which taxonomic and historical information is embedded within specimens, labels and their physical arrangement. However, in large entomological collections, much of this information remains inaccessible because it is not captured in standardised digital formats. Whole-drawer imaging offers a scalable approach to digitisation, but existing workflows primarily focus on specimen-level extraction and often neglect the curatorial structure encoded in drawer organisation. Here, we present an open-source workflow that transforms whole-drawer images into structured, machine-readable inventory data, while preserving their spatial arrangement and taxonomic context. The pipeline integrates low-cost imaging, barcode linkage, deep learning-based object detection (YOLOv.11) and optical character recognition (OCR) to reconstruct taxonomic series and estimate specimen counts per taxon from drawer images. The approach was developed for the Coleoptera collection of the Senckenberg Research Institute Frankfurt, encompassing approximately 1.9 million specimens which are stored in more than 5,000 drawers. Model training on Carabidae collection drawers overall achieved high performance (precision 99.2%, recall 97.9%, mAP0.5 98.5%), with reasonable transferability to other beetle families and Hymenoptera drawers tested. By treating the drawer as the primary unit of digitisation, this workflow provides a scalable intermediate layer between physical collections and specimen-level databases. It enables rapid assessment of taxonomic composition, supports collection management and digitisation planning and contributes to the mobilisation of biodiversity data from large historical collections.

Keywords

Deep learning, digitisation, entomological collection, natural history museums, object detection, OCR

Fulltext

From pixels to data: automated reconstruction of taxonomic series from whole-drawer images in historical insect collections