How konstantinkozhin/manuscript-ocr works
Manuscript OCR is a Python framework for OCR/HTR on manuscript and modern document images. Its primary workflow accepts an image, detects text regions, orders them, recognizes text, and can optionally correct the result, represented as structured page data. The repository also contains model-training components for recognition, detection, and text correction. The pipeline excerpt establishes image loading and default detector/layout choices, but omits most stage execution; those stage links are therefore shown as unverified rather than asserted. README capability statements support the broader OCR workflow.
Main entry points: _pipeline.py OCR pipeline · structures.py Page results · __init__.py Text recognition · base.py Model artifacts
Components 21
OCR workflow
- OCR pipeline
src/manuscript/_pipeline.py - Text detection
src/manuscript/detectors/_east/__init__.py - Reading order
src/manuscript/layouts/_simplesorting/__init__.py - Region preparation
src/manuscript/recognizers/_common/region_preparers.py - Text recognition
src/manuscript/recognizers/_trba/__init__.py - Text correction
src/manuscript/correctors/_charlm/__init__.py
Shared runtime
- Page results
src/manuscript/data/structures.py - Image I/O
src/manuscript/utils/io.py - Image geometry
src/manuscript/utils/geometry.py - Model artifacts
src/manuscript/api/base.py
Inference models
- YOLO detector
src/manuscript/detectors/_yolo/__init__.py - EAST detector
src/manuscript/detectors/_east/__init__.py - TRBA network
src/manuscript/recognizers/_trba/model/model.py - Correction model
src/manuscript/correctors/_charlm/model.py
Model training
- Recognition data
src/manuscript/recognizers/_trba/data/dataset.py - Recognition transforms
src/manuscript/recognizers/_trba/data/transforms.py - Recognition trainer
src/manuscript/recognizers/_trba/training/train.py - Detection data
src/manuscript/detectors/_east/dataset.py - Detection training
src/manuscript/detectors/_east/train_utils.py - Correction trainer
src/manuscript/correctors/_charlm/train.py
Other components
- Image user
Connections 23
- Image user → OCR pipeline: submits image
- OCR pipeline → Image I/O: loads image
- OCR pipeline → Text detection: runs detection
- OCR pipeline → Reading order: orders regions
- OCR pipeline → Text recognition: runs recognition
- OCR pipeline → Text correction: may correct
- Text detection → Page results: produces regions
- Reading order → Page results: orders page data
- Text recognition → Region preparation: prepares crops
src/manuscript/recognizers/_trba/__init__.py - Region preparation → Image geometry: crops regions
src/manuscript/recognizers/_common/region_preparers.py - Text recognition → Page results: updates text
- Text correction → Page results: corrects text
src/manuscript/correctors/_charlm/__init__.py - OCR pipeline → Page results: returns result
- OCR pipeline → YOLO detector: defaults to
src/manuscript/_pipeline.py - Text detection → Model artifacts: loads weights
src/manuscript/detectors/_east/__init__.py - Text recognition → Model artifacts: loads weights
src/manuscript/recognizers/_trba/__init__.py - Text correction → Model artifacts: loads weights
src/manuscript/correctors/_charlm/__init__.py - Detection training → Detection data: uses dataset
- Recognition trainer → Recognition data: uses dataset
src/manuscript/recognizers/_trba/training/train.py - Recognition trainer → Recognition transforms: uses transforms
src/manuscript/recognizers/_trba/training/train.py - Recognition trainer → TRBA network: trains model
src/manuscript/recognizers/_trba/training/train.py - Correction trainer → Correction model: trains model
src/manuscript/correctors/_charlm/train.py - EAST detector → Model artifacts: loads weights
src/manuscript/detectors/_east/__init__.py