GOST-OCR

Automatic stamp localization and metadata extraction from GOST technical drawings.

Scanned archival drawings contain all project metadata — document number, author, revision, date — locked inside a title block. This tool finds that block and extracts the text, turning a folder of scans into a searchable catalog without manual re-entry.

Built as the first stage of a scan-to-BIM pipeline for archival engineering documentation.

---

Results

Evaluated on 49 real archival drawings (DWG exports, various DPI, mixed quality):

Method	Mean IoU	Median IoU	Successful (IoU > 0.5)	Wins
OpenCV (auto ROI)	0.655	0.962	31 / 49	38
YOLO (synthetic training)	0.351	0.207	16 / 49	10

Key finding: OpenCV with automatic ROI selection based on document orientation significantly outperforms YOLO trained on synthetic data. GOST standardization of stamp position and dimensions makes geometric heuristics more robust than a neural network trained on 25 images.

YOLO failure analysis: the model detects all rectangular stamp-like elements, not just the main title block. Two-class training (main stamp vs. auxiliary elements) is the recommended next step.

See docs/article/ARTICLE.md for the full comparison paper.

---

How It Works

Three-stage pipeline:

Scan → Preprocess → Localize → Extract
        (deskew,     (OpenCV     (EasyOCR)
         auto ROI)    or YOLO)

Preprocessing — deskew correction, automatic DPI detection, ROI selection based on page orientation (landscape → bottom-right, portrait → bottom, tall portrait → right).

Localization — contour-based detection filtered by GOST 2.104 aspect ratios (FORM_3: 3.36, FORM_4: 1.61, FORM_5: 5.40), or YOLOv8n trained on synthetic stamps.

Extraction — EasyOCR over the localized region, structured JSON output with per-field bounding boxes and confidence scores.

---

Installation

Requires Python 3.12+.

git clone https://github.com/Sapar-hub/gost-ocr.git
cd gost-ocr
pip install uv
uv pip sync pyproject.toml

---

Usage

# Full pipeline (preprocess + localize + OCR)
uv run gost-ocr pipeline /path/to/drawings/ --detector=auto --debug

# Localization only
uv run gost-ocr localize /path/to/drawings/ --detector=opencv

# Reproduce benchmark results
uv run python -m src.gost_ocr.benchmark.compare

Detection methods: auto (YOLO → OpenCV fallback), opencv, yolo

ROI options: auto (default), bottom_right, bottom, right, full_page, and others

Output: JSON per image with stamp_bbox, text_blocks (text + confidence + coordinates), full_text

{
    "stamp_bbox": [1222, 1091, 567, 167],
    "text_blocks": [
        {"text": "Лист", "confidence": 0.999, "box": [[29,15],[83,15],[83,38],[29,38]]}
    ],
    "full_text": "Лист МГТ-2024-ПЗ..."
}

---

Dataset

• Training: 25 synthetic images generated per GOST 2.104 (DPI 200/300/400, FORM_3/4/5)
• Test: 49 real drawings from open DWG archives, manually labeled with LabelImg
• Ground truth: YOLO format, src/gost_ocr/datasets/labels/test/
• Model weights: src/gost_ocr/models/yolo/best.pt

---

Roadmap

• Two-class YOLO training: main stamp vs. auxiliary elements
• Field-level classification (document number, author, date, revision) per GOST 2.104 cell layout
• Dimension chain recognition (размерные цепочки) for geometry extraction
• IFC/JSON export for BIM integration

---

Stack

Python 3.12 · OpenCV · YOLOv8n · EasyOCR · Typer · uv

---

Paper

Сравнение методов локализации штампов технической документации ГОСТ: OpenCV vs YOLO