Drawing Bubble Detection

Offline CPU-based engineering drawing bubble and dimension detection system.

This project detects numbered balloons/bubbles in engineering drawings, traces leader lines, reads dimension callouts using OCR, and returns bubble-to-dimension assignments with annotated result images and live pipeline diagnostics.

Current Milestone

This repository is frozen as a project milestone for showcasing the work completed so far.

Validated showcase results:

Case	Bubble Recall	Value Accuracy	Notes
ui_screenshot_1png	100%	100%	Clean UI screenshot
dense_drawing_3png	100%	100%	Dense drawing with rotated/local OCR recovery
mn_m1	100%	100%	Clean stress-suite image
mn_n1	100%	100%	Clean stress-suite image

The system is strongest on clean CAD exports, clean screenshots, and selected dense drawings. Photo/monitor captures remain the main production challenge.

Screenshots

Pipeline execution with live progress, per-step timings, and request logs:

Annotated detection result with bubble-to-dimension assignments:

Clean drawing result with detected balloons and assigned dimension values:

Tech Stack

• Python
• FastAPI
• OpenCV
• RapidOCR
• ONNX Runtime
• NumPy / SciPy
• ASP.NET Core MVC frontend
• REST API integration

What The System Does

• Detects colored and circular engineering drawing bubbles.
• Reads bubble numbers using OCR and local recovery.
• Groups OCR text into engineering callouts such as dimensions, tolerances, chamfers, radius/diameter values, and notes.
• Traces leader lines from bubbles to dimension text.
• Assigns bubble numbers to the most likely dimension/callout.
• Produces annotated result images.
• Shows live pipeline logs and step timings in the ASP.NET UI.
• Generates benchmark reports for scored validation cases.

Important Files

Path	Purpose
DrawingBubbleService_M1/detector.py	Main detection pipeline
DrawingBubbleService_M1/main.py	FastAPI service
DrawingBubbleService_M1/benchmark_production.py	Production benchmark runner
DrawingBubbleService_M1/benchmark_cases.json	Benchmark manifest
DrawingBubbleService_M1/BENCHMARKING.md	Benchmarking workflow
DrawingBubbleService_M1/ocr_rules.py	OCR normalization and token classification
DrawingBubbleService_M1/callout_rules.py	Dimension/callout grouping logic
DrawingBubbleService_M1/leader_geometry.py	Leader-line geometry helpers
Controllers/DrawingBubbleController.cs	ASP.NET controller integration
Views/DrawingBubble/*.cshtml	ASP.NET UI pages

Running The Python Service

cd DrawingBubbleService_M1
python -m venv .venv
.\.venv\Scripts\activate
pip install -r requirements.txt
copy .env.example .env
python main.py

The API runs locally and is designed to work offline.

Running Benchmarks

cd DrawingBubbleService_M1
python benchmark_production.py --cases ui_screenshot_1png dense_drawing_3png mn_m1 mn_n1 --skip-unified --fail-on-threshold
pytest test_benchmark_manifest.py -q

Benchmark outputs include:

• benchmark_report.json
• benchmark_report.csv
• benchmark_readiness.md
• Annotated result images

Generated benchmark/debug outputs are intentionally ignored by Git.

Demo Assets

Demo recordings and result logs are listed here:

docs/demo.md

Current Limitations

The current implementation is mostly rule-based with OCR and geometry heuristics. It works well on clean images but is not yet fully production-ready for all unseen images.

Known difficult cases:

• Phone/monitor photos
• Perspective distortion
• Low contrast or noisy scans
• Very dense drawings
• Missing/partial leader lines
• Rotated or vertical dimension text
• OCR confusion between similar characters or digits

For production deployment, wrong assignments should be treated carefully. Low-confidence results are marked for review where possible.

Recommended Production Path

To make this production-ready for unseen customer drawings:

1. Build a larger labelled dataset of real engineering drawings.
2. Label bubbles, leader lines, arrowheads, and dimension text regions.
3. Train or fine-tune an offline detector for drawing regions of interest.
4. Use targeted OCR only on detected callout regions instead of relying on full-page OCR.
5. Keep confidence-based validation so uncertain results are sent for manual review instead of forced.
6. Expand benchmark coverage and require regression tests before every change.

Repository Status

This repository contains the final milestone code and documentation for showcasing the work completed on the offline engineering drawing bubble and dimension detection system.