An open-source framework for integrating multi-source layout and text recognition tools into scalable OCR workflows
Kay-Michael Würzner
Konstantin Baierer
Bibliotheca Baltica 2018 Rostock, 2018-10-05
https://ocr-d.github.io/2018-10-05-baltica/
https://ocr-d.github.io/2018-10-05-baltica/slides.pdf
Overview
- Why OCR-D
- The OCR-D initiative
- Architecture
- State of OCR-D tools
- Scalability
- Open source
Users
want
text data
massive amounts
structured
easily accessible
Libraries
provide
text data
large amounts
unstructured
hard to access
Why the discrepancy?
Underspecified requirements on OCR by funders and users
OCR of historical data of little economic interest
Little competition
Academical solutions non-sustainable
Inflexible workflows
The OCR-D initiative
DFG-organized expert Workshop (2014)
Verfahren zur Verbesserung von OCR-Ergebnissen
Result: A concerted effort for improving OCR is seen as required.
OCR-D Coordination project
 |
 |
 |
 |
Phase 1: Exploring the domain (2015-2017)
Surveyed (open-source) ecosystem
around OCR and OLR
Identified Tasks
Prepared call for proposals for DFG
Phase 2: Module project stage (2018-2019)
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Phase 3: Going productive (2018-2020)
Integrate with existing digitization workflow software, e.g. Kitodo
Make OCR-D-developed software uniformly deployable
Advise DFG on OCR requirements for "Praxisrichtlinien"
Not in this talk (but in OCR-D)
Ground Truth for OCR
OCR engine training
OCR research data repository
Workflow composition and provenance
Long-time preservation of OCR
Architecture
"Multi source"
existing tools by OCR-D partners
(tesseract, PoCoTo, LAREX...)
new developments within OCR-D
(font identification, post-correction...)
existing tools outside OCR-D
(ocropus, kraken, ScanTailor, OLENA...)
😍 Modular 😍
😭 Monolithic 😭
Specification
+
Implementation
OCR-D/spec
METS + PAGE-XML (+ ALTO)
Structured tool descriptions
Command line interface
HTTP interface
Actionable documentation
OCR-D/core
Validation and helper functions
Python library
Shell library
Why Python?
Python widely used in computer vision and machine learning
(keras, pytorch...)
Wrapping existing tools with minimal friction
(ocropus, kraken ...)
Bindings for low-level APIs
(opencv, tesserocr ...)
Why shell?
Lowest common denominator
Wrap arbitrary command line tools
Process callout possible in every
framework/workflow engine/programming environment
State of the OCR-D toolset
Preprocessing
| Tool | Developer | Functionality | Wrapper |
|---|---|---|---|
| anyOCR | DFKI Kaiserslautern | binarization, cropping, deskewing, dewarping | (python) |
| OLENA | OCR-D | binarization | shell |
| tesseract | UB Mannheim, ASV Leipzig |
binarization | python |
| OCRopus | OCR-D | binarization | python |
| kraken | OCR-D | binarization | python |
| ImageMagick | OCR-D | binarization, conversion | shell |
Layout recognition
| Tool | Developer | Functionality | Wrapper |
|---|---|---|---|
| anyOCR | DFKI Kaiserslautern | block+line seg8n, block class7n, document analysis | (python) |
| LAREX | Uni Würzburg | block+line seg8n, block class7n | (shell) |
| OCRopus | OCR-D | line seg8n | python |
| kraken | OCR-D | line seg8n | python |
| tesseract | UB Mannheim, ASV Leipzig |
block+line seg8n | python |
| dh_segment | OCR-D | block+line seg8n | (shell) |
Text recognition
| Tool | Developer | Functionality | Wrapper |
|---|---|---|---|
| OCRopus | OCR-D | text recognition | python |
| kraken | OCR-D | text recognition | python |
| tesseract | UB Mannheim, ASV Leipzig |
text recognition | python |
| calamari | OCR-D | text recognition | (python) |
| ocrad | OCR-D | text recognition | (shell) |
Postprocessing
| Tool | Developer | Functionality | Wrapper |
|---|---|---|---|
| corASV | ASV Leipzig | post correction | (python) |
| PoCoTo | CIS München | post correction | python |
| keraslm | ASV Leipzig | post correction | python |
| ocrevalUAtion | OCR-D | evaluation | (shell) |
Your tool?
Scalability
<IMPRESSIVE NUMBER HERE>
Geared towards real digitization scenarios
Cooperation with Kitodo and commercial providers
Frequent reality check with current practices ("Pilotbibliotheken")
Different deployment options
Modularity + uniform interfaces
⇒
Adaptive workflows
(Instantiation and composition up to users)
Open Source is more than "Open Source"
Step 1: Get funded!
Step 2: Develop!
Step 3: Publish code!
Sustainability and reuse!
Best Practices
Transparency from day one
Unit tests
Unified test assets
Continuous Integration
Semantic versioning
Docker base image
Releases to GitHub, PyPI, DockerHub
Community
Issues
Pull requests
Code review
Support chat