dotmark-qa

Edge Vision QA System for Tyre Paint Mark Inspection

A production-ready, demo-optimized quality assurance system for detecting and analyzing paint marks (dots, donuts, multiple marks) on tyre sidewalls. Built for edge deployment with fast iteration, reliability, and clarity.

Features

Vision Pipeline: Color-based detection (HSV) + contour analysis for red/yellow paint marks
Shape Detection: Dot (filled circle), donut (ring), and multiple marks
Quality Metrics: Fill ratio, circularity, edge continuity, overspray, ring continuity
Profile System: SKU-based configuration with hot-reload
OCR Integration: Optional sidewall text reading for automatic SKU detection
Real-time HMI: React TypeScript interface with live preview and overlay
Mock Components: Camera simulator and PLC webhook for demo
Fast Processing: Target p95 latency ≤150ms
Complete Testing: pytest suite with profile, analyzer, and decision tests

Architecture

┌─────────────────┐     WebSocket      ┌──────────────┐
│   React HMI     │ ←─────────────────→ │              │
│   (Port 5173)   │                     │              │
│                 │     REST API        │  FastAPI     │
│  - Run View     │ ←─────────────────→ │  Edge        │
│  - Review       │                     │  Service     │
│  - Tuning       │                     │  (Port 8080) │
│  - System       │                     │              │
└─────────────────┘                     │  - Vision    │
                                        │  - OCR       │
                                        │  - Storage   │
                                        └──────┬───────┘
                                               │
                                               │ Webhook
                                               ↓
                                        ┌──────────────┐
                                        │  Mock PLC    │
                                        │  (Port 9090) │
                                        └──────────────┘

Tech Stack

Backend (Edge Service):

Python 3.10+
FastAPI + uvicorn
OpenCV, NumPy, scikit-image
SQLModel (SQLite)
Tesseract OCR (optional)
orjson for logging

Frontend (HMI):

React 18 + TypeScript
Vite build tool
WebSocket for live preview
Canvas overlay rendering

Infrastructure:

Docker + docker-compose
pytest for backend tests
Mock camera & PLC for demo

Quick Start

Option 1: Docker (Recommended)

# Generate sample images
make samples

# Start all services
make docker-up

# Access HMI
open http://localhost:5173

Services:

HMI: http://localhost:5173
API Docs: http://localhost:8080/docs
Mock PLC: http://localhost:9090/health

Option 2: Local Development

Terminal 1 - Edge Service:

cd apps/edge
pip install -r requirements.txt
python scripts/generate_samples.py  # Generate test images
python scripts/seed_db.py          # Initialize database
python main.py

Terminal 2 - HMI:

cd apps/hmi
npm install
npm run dev

Terminal 3 - Mock PLC:

cd apps/mock-plc
npm install
npm start

Access HMI at http://localhost:5173

Repository Structure

dotmark-qa/
├── apps/
│   ├── edge/                   # FastAPI edge service
│   │   ├── api/               # API routes
│   │   ├── core/              # Config, logging, schemas
│   │   ├── vision/            # Image analysis pipeline
│   │   ├── ocr/               # OCR profile resolver
│   │   ├── storage/           # Database models
│   │   ├── profiles/          # Profile loader
│   │   ├── mocks/             # Camera simulator
│   │   ├── tests/             # pytest tests
│   │   └── main.py
│   ├── hmi/                   # React TypeScript HMI
│   │   ├── src/
│   │   │   ├── components/
│   │   │   ├── views/         # Run, Review, Tuning, System
│   │   │   ├── services/      # API client
│   │   │   └── main.tsx
│   │   └── Dockerfile
│   └── mock-plc/              # Webhook sink
│       └── server.js
├── configs/
│   ├── camera.yaml            # Camera & ROI config
│   └── profiles.yaml          # Quality profiles
├── data/
│   └── samples/               # Test images (generated)
├── scripts/
│   ├── generate_samples.py    # Create test images
│   ├── seed_db.py            # Initialize database
│   └── run_local.sh          # Local run helper
├── docker-compose.yml
├── Makefile
├── README.md
└── LICENSE

API Overview

POST /api/inspect

Perform inspection.

Request:

{
  "sku": "215/60R16",           // Optional
  "profile_id": "yellow_single_dot",  // Optional
  "capture": true,              // Capture from camera
  "image_b64": null             // Or provide base64 image
}

Response:

{
  "ts": "2025-01-15T10:30:00Z",
  "sku": "215/60R16",
  "profile_id": "yellow_single_dot",
  "result": "PASS",
  "reasons": [],
  "mark_count": 1,
  "per_mark": [
    {
      "color": "yellow",
      "shape": "dot",
      "fill": 0.92,
      "circ": 0.88,
      "edge_cont": 0.94,
      "overspray": 0.02,
      "bbox": [720, 360, 90, 90]
    }
  ],
  "latency_ms": 127.5,
  "image_path": "data/captures/20250115_103000_000.png",
  "overlay_path": "static/overlays/20250115_103000_000_overlay.png"
}

GET /api/profiles

List all profiles.

GET /api/last

Get last inspection result.

GET /api/stats?sku=&result=

Get statistics.

WS /api/ws/preview

Live preview stream (JPEG frames as base64).

Quality Profiles

Profiles define expected marks per SKU. See configs/profiles.yaml:

profiles:
  - profile_id: "yellow_single_dot"
    sku_regex: "^(215/60R16|225/55R17).*"
    allowed_colors: [yellow]
    shapes: [dot]
    count_min: 1
    count_max: 1
    allow_zero: false
    thresholds:
      dot:
        fill_min: 0.85
        circularity_min: 0.80
        edge_cont_min: 0.92
        overspray_max: 0.05
      color:
        yellow_hue: [40, 75]
        sat_min: 0.35
        val_min: 0.40

Thresholds:

Dot:
- fill_min: Minimum fill ratio vs fitted circle
- circularity_min: 4πA/P² (1.0 = perfect circle)
- edge_cont_min: Edge continuity (arc coverage)
- overspray_max: Maximum overspray outside ideal circle
Donut:
- ring_cont_min: Ring continuity (perimeter coverage)
- thickness_cov_max: Coefficient of variation for ring thickness
Color:
- yellow_hue: HSV hue range [40, 75]
- red_hue_low/high: Red wraps hue circle
- sat_min: Minimum saturation
- val_min: Minimum value (brightness)

Vision Pipeline

Preprocess: Crop ROI, denoise (fastNlMeans), optional CLAHE
Color Extraction: HSV masks for yellow/red with morphology
Contour Detection: Find connected components, filter by area
Shape Classification: Detect inner contours for donut vs dot
Metrics Computation:
- Dot: Fill, circularity, edge continuity, overspray
- Donut: Ring continuity, thickness CoV
Decision: Evaluate against profile thresholds
Overlay Rendering: Annotate with metrics and results

Reasons:

low_fill, poor_circularity, edge_gap, smear (overspray)
wrong_count, wrong_color, unexpected_mark
spec_unknown (no profile found)
faded_low_snr

Testing

# Run backend tests
cd apps/edge
pytest

# Or via Makefile
make test

Tests cover:

Profile loading and SKU resolution
Vision analyzer (color masks, contours, metrics)
Decision engine (thresholds, count, allow_zero)

Demo Script (Stakeholder Meeting)

Setup (5 min before meeting)

make samples
make docker-up
# Wait for services to start (~30s)

Live Demo (10 minutes)

1. Introduction (1 min)

Open http://localhost:5173
Show clean HMI with 4 tabs: Run, Review, Tuning, System

2. Run View - Manual Inspection (3 min)

Select SKU: Choose "215/60R16" from dropdown
Click "Inspect Now"
Observe:
- Live preview updates
- Overlay appears with green contours
- Result badge shows PASS (green)
- Metrics: SKU, Profile, Marks Found: 1, Latency: ~120ms
- Per-mark details: fill=0.92, circ=0.88, edge=0.94
- Recent results table updates

3. Auto Trigger Mode (2 min)

Enable "Auto Trigger (3s)" checkbox
Watch: Table populates every 3 seconds with mix of PASS/FAIL
Point out:
- Some failures appear (red badges)
- Reasons displayed: "low_fill", "edge_gap", "smear"
- Latency stays under 150ms

4. Failure Analysis (2 min)

Disable auto trigger
Manually click through until you get a FAIL
Show overlay: Red annotations on problematic areas
Highlight reasons: e.g., "edge_gap" → missing arc segment
Show Mock PLC terminal: Failure logged with details

5. System View (1 min)

Click "System" tab
Show:
- Health: Camera: mock, Profiles Loaded: 4
- Statistics: Pass rate, Avg latency
- Top failure reasons bar chart
- Link to API docs

6. Tuning View (1 min)

Click "Tuning" tab
Select profile: "yellow_single_dot"
Show thresholds: JSON view of fill_min, circ_min, etc.
Explain: "Operators can adjust these via YAML (hot-reload enabled)"

Key Messages:

✅ Fast: <150ms latency
✅ Reliable: Configurable thresholds per SKU
✅ Transparent: Full metrics and overlay for every inspection
✅ Production-ready: Mock components swap for real camera/PLC
✅ Extensible: Vision pipeline ready for ML upgrade (UNet/YOLO)

Configuration

Camera (configs/camera.yaml)

camera:
  type: mock  # mock | basler | ids
  width: 2048
  height: 1536
  fps: 30
  roi:
    enabled: true
    x: 256
    y: 384
    width: 1536
    height: 768

Mock Camera

Loads images from data/samples/ in round-robin
Optional jitter (brightness, hue) for variability
Replace with real camera driver (Basler/IDS) in production

Mock PLC

Receives inspection results via webhook
Logs to console with color-coded PASS/FAIL
Replace with real PLC I/O (Profinet/EtherNet/IP) in production

Roadmap

Phase 2: ML Upgrade

Swap color/contour pipeline with UNet/YOLO-Seg
Train on real tyre data
Add confidence scores
Handle occlusions and complex backgrounds

Phase 3: Production Deployment

Real camera integration (Basler/IDS)
PLC I/O (Profinet, EtherNet/IP)
Encoder tracking for trigger timing
Multi-camera setup
Authentication & role-based access

Phase 4: Advanced Features

Defect classification beyond marks (cracks, bubbles)
Statistical process control (SPC) charts
Remote monitoring dashboard
Model versioning and A/B testing

Performance

Latency (measured on sample images):

p50: ~100ms
p95: ~130ms
p99: ~145ms

Target: p95 ≤150ms on 5–12 MP downscaled ROI

Troubleshooting

Services won't start:

docker-compose logs edge

No sample images:

python3 scripts/generate_samples.py

Tests fail:

cd apps/edge
pip install -r requirements.txt
pytest -v

HMI shows "No preview":

Check edge service is running: curl http://localhost:8080/api/health
Check WebSocket connection in browser console

License

MIT License - See LICENSE

Contributing

This is a demo repository. For production use:

Replace mock camera with real camera driver
Integrate actual PLC I/O
Tune thresholds on real production data
Consider ML-based segmentation for complex scenarios
Add authentication and audit logging

Support

For questions or issues, contact the engineering team or file an issue in this repository.

Built with ❤️ for reliable edge vision QA

Name		Name	Last commit message	Last commit date
Latest commit History 1 Commit
apps		apps
configs		configs
data/samples		data/samples
scripts		scripts
.gitignore		.gitignore
DEPLOYMENT.md		DEPLOYMENT.md
LICENSE		LICENSE
Makefile		Makefile
QUICKSTART.md		QUICKSTART.md
README.md		README.md
STRUCTURE.txt		STRUCTURE.txt
docker-compose.yml		docker-compose.yml

License

replit4tj/apollo-vision-qa-demo

Folders and files

Latest commit

History

Repository files navigation

dotmark-qa

Features

Architecture

Tech Stack

Quick Start

Option 1: Docker (Recommended)

Option 2: Local Development

Repository Structure

API Overview

POST /api/inspect

GET /api/profiles

GET /api/last

GET /api/stats?sku=&result=

WS /api/ws/preview

Quality Profiles

Vision Pipeline

Testing

Demo Script (Stakeholder Meeting)

Setup (5 min before meeting)

Live Demo (10 minutes)

Configuration

Camera (configs/camera.yaml)

Mock Camera

Mock PLC

Roadmap

Phase 2: ML Upgrade

Phase 3: Production Deployment

Phase 4: Advanced Features

Performance

Troubleshooting

License

Contributing

Support

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Languages

Packages