A comprehensive review of detection, optical character recognition, and structuring methods for herbarium label digitization

Global herbarium digitization has produced over 50 million specimen images, many of which contain handwritten or printed labels with rich metadata such as species names, collection dates, localities, and ecological notes. Manual transcription of this information is labor intensive, error prone, and not scalable, which has motivated the development of automated solutions. This systematic review examines studies published between 2018 and 2026, identified through electronic database searches and backward reference screening, focusing on deep learning and computer vision approaches for herbarium label digitization. The reviewed methods include convolutional neural networks, transformer-based models, large language models, and open-source pipelines for optical character recognition and structured metadata extraction from historical and handwritten botanical records. The results highlight significant advances in label detection, text recognition accuracy, and metadata structuring enabled by deep learning techniques, although reported performance varies considerably depending on dataset quality, handwriting complexity, and domain adaptation strategies. Despite this progress, substantial challenges remain in handling historical handwriting styles, heterogeneous label layouts, and multilingual collections. Future research should emphasize the creation of curated benchmark datasets and the development of scalable, robust artificial intelligence (AI) pipelines to enhance the reliability, efficiency, and generalizability of automated herbarium digitization systems.