<Phi2AI/>
HTR: Making Manuscripts Searchable
Computer Vision

HTR: Making Manuscripts Searchable

#Computer-Vision#eScriptorium#Heritage#IIIF#Europeana#Open-Source
Computer Vision // Pipeline

  • Unlock the Black Box: Converting static manuscript images into fully searchable text using custom HTR (Handwritten Text Recognition).

  • Local Infrastructure: Implemented eScriptorium with Kraken locally to train on specific handwriting styles without cloud dependency.

  • Interoperability: Output complies with IIIF standards, allowing seamless integration with the global Cultural Heritage network.

The Problem: The “Black Box” of Archives

Digitization is not enough. Taking a photo of a manuscript preserves the image, but the content remains locked inside. You cannot Ctrl+C a text from JPEG of a handwritten diary.

For the Milutin Bojić Library, we faced a challenge: thousands of pages of the poet’s original manuscripts were digitized but textually inaccessible. Standard OCR failed completely on his unique cursive script.

The Solution: Custom HTR with eScriptorium

We implemented a full Handwritten Text Recognition (HTR) pipeline using a local instance of eScriptorium. Unlike commercial cloud solutions, this allowed us to train a model specifically for Bojić’s handwriting style without data leaving our local server.

The Workflow

  1. Segmentation: We used the Kraken engine to analyze page layouts, separating text bodies from marginalia.
  2. Ground Truth: We manually transcribed a small subset (~50 pages) to serve as the “textbook” for the AI.
  3. Training: We fine-tuned a Kraken model on our Ground Truth data.
  4. Inference: The model achieved a Character Error Rate (CER) low enough for effective full-text search.

Interoperability: IIIF & Europeana

This project wasn’t just about reading text; it was about sharing it. By implementing the IIIF (International Image Interoperability Framework) standard, our local server acts as a node in the global digital heritage network.

Global Visibility

Because our metadata is standardized, the entire handwritten collection is automatically harvested and visible on the Europeana Portal, making Serbian heritage accessible to researchers worldwide.

Interactive Manuscript Search

This instance loads the live IIIF Manifest. We have pre-populated the search query видећи to demonstrate the HTR capability.

Artifact Resources

Back to Artifacts