Automated information extraction from plant specimen labels using OCR and large language models.

Area of Science:

• Botany
• Biodiversity Informatics
• Digital Curation

Background:

• Digitizing herbarium specimens is vital for biodiversity research and data sharing.
• Manual transcription is inefficient and faces challenges with large volumes, varied layouts, and handwriting.
• Existing methods struggle with the complexity of specimen labels.

Purpose of the Study:

• To develop an automated pipeline for structured information extraction from herbarium specimen labels.
• To overcome bottlenecks in manual data entry and improve digitization efficiency.
• To extract 16 key metadata fields from both printed and handwritten labels.

Main Methods:

• Integration of the PaddleOCR engine with the DeepSeek large language model (LLM).
• Development of a pipeline for structured information extraction from specimen labels.
• Implementation of a confidence-based quality control mechanism for handwritten labels.

Main Results:

• Achieved 95.4% field-level accuracy for printed labels.
• The LLM effectively corrected upstream OCR errors, showing a weak correlation between OCR confidence and final accuracy (r = 0.32).
• The system maintained functionality for handwritten labels, acknowledging its limitations.

Conclusions:

• The proposed hybrid pipeline offers a robust, scalable, and practical solution for accelerating botanical collection digitization.
• Local image processing ensures data security, and text-only LLM parsing enhances cost-efficiency.
• This method significantly improves biodiversity data creation and sharing efficiency for real-world applications.