Technical Drawing Vectorization with OpenCV and SVG Export

中文版

An OpenCV portfolio project for technical drawing cleanup, geometry detection, and SVG vectorization.

Pipeline

PNG/JPEG/PDF page image → preprocessing → geometry detection → primitive normalization → SVG/JSON export → overlay & report

Quick Start

# Install dependencies
uv sync --extra dev

# Generate synthetic test fixtures
uv run tdv-make-fixtures -o data/fixtures/synthetic

# Vectorize a single image
uv run tdv-vectorize data/fixtures/synthetic/composite.png -o data/results/runs/my-run

# Batch process a directory
uv run tdv-vectorize data/fixtures/synthetic -o data/results/runs/batch

# Evaluate against ground truth
uv run tdv-evaluate data/fixtures/synthetic -o data/results/runs/eval

# All commands require no external API keys

Output

Asset	Format	Description
SVG	.svg	Vector rendering with color-coded layers (lines, circles, arcs, polylines)
Primitives	.json	Machine-readable structured geometry
Overlay	.png	Detection drawn over cleaned image
Stages	.png	Every preprocessing step saved for comparison
Report	.json/.md	Per-fixture Precision/Recall/F1 metrics

Configuration

See configs/default.yaml. All parameters are typed via pydantic and can be overridden at runtime:

uv run tdv-vectorize input.png -c my_config.yaml -o results/

Capabilities

• Preprocessing: grayscale, denoise (fastNlMeans/bilateral), CLAHE contrast, adaptive/OTSU threshold, Hough/minAreaRect deskew, contour-based perspective correction
• Geometry detection: probabilistic Hough lines, Hough circles, contour-based arcs, polygon approximation
• Normalization: collinear line merging, endpoint snapping, length-based noise filter
• Export: SVG with <g> layers, JSON primitives, DXF via ezdxf
• PDF input: via pypdfium2 (no system poppler)
• Deterministic: same input + config → byte-identical JSON and SVG

Known Limitations

• Hough parameters require per-image tuning for optimal detection; defaults work best on clean line drawings
• Deskew/perspective may misalign clean synthetic images (designed for real scanned/photographed drawings)
• Arc detection is heuristic (contour min-area-rect fit); accuracy varies
• DXF export is limited to basic entities (lines, circles, arcs, polylines) — no blocks, attributes, or advanced features
• No deep learning, LLM, or OCR integration (by design)
• Pixel IoU metric assumes binary-cleaned image; noisy backgrounds affect accuracy

No External API Key Required

All processing is local. No cloud services, no API keys, no internet connection needed.

License

MIT