Lesson 2 of 5·10 min read

Quality Control with Vision AI

In manufacturing, quality defects cost 3–5% of annual revenue — through scrap, rework, complaints, and reputation damage. Human inspectors get fatigued, are subjective, and can't work 24/7. Vision AI changes that.

Defect Detection in Manufacturing

Typical Defects

Vision AI detects flaws that the human eye often misses:

  • Surface defects: Scratches, dents, discoloration, cracks
  • Dimensional deviations: Wrong dimensions, warping, asymmetry
  • Assembly errors: Missing parts, wrong orientation, loose connections
  • Material defects: Inclusions, voids, porosity
  • Coating defects: Bubbles, peeling, uneven layer thickness

Camera Setup

The right hardware is crucial:

TechnologyUse CaseAdvantage
2D camerasSurface inspectionAffordable, fast
3D camerasDimension checkingDetects heights/depths
InfraredThermal defectsSees below surfaces
HyperspectralMaterial analysisDetects chemical differences
X-ray/CTInternal defectsSee through material

Training with Limited Data

Defects are rare — sometimes only 10–50 images available. Solutions:

  • Data augmentation: Rotation, mirroring, color variation, cropping
  • Synthetic data: Artificially render defects onto good parts
  • Few-shot learning: Models that learn from few examples (e.g., PatchCore)
  • Anomaly detection: Only train on "good" — everything else is a defect

ROI Calculation

Cost of a Vision AI Solution

Typical investments for one production line:

  • Hardware: €15,000–50,000 (cameras, lighting, edge computer)
  • Software: €20,000–80,000 (license or custom development)
  • Integration: €10,000–30,000 (PLC connection, data integration)
  • Running costs: €5,000–15,000/year (maintenance, updates, cloud)

Total: €50,000–175,000 for one line — typically pays for itself in 6–18 months.

Savings

  • Scrap reduction: 30–70% fewer defects pass final inspection
  • Rework costs: −50% through early defect detection
  • Personnel costs: 1 vision system replaces 2–4 manual inspection stations (shift operation)
  • Complaint costs: −40% through higher delivery quality
  • Downtime: −20% through predictive maintenance (defect patterns as early indicator for machine problems)

Case Study

An automotive supplier with 500,000 parts/month:

  • Before: 2.3% scrap rate, 4 inspectors per shift, €180,000/year complaint costs
  • After: 0.4% scrap rate, 1 inspector per shift (monitoring), €45,000/year complaint costs
  • ROI: Investment €120,000 — savings €280,000/year → payback in 5 months

Best Practices

  1. Lighting is 80% of success — invest more in light than in cameras
  2. Define golden samples: Reference images for "good" and each defect class
  3. Edge-first: Processing directly at the line, not in the cloud (latency!)
  4. Feedback loop: Inspectors validate AI decisions → model improves
  5. Start with the worst: Begin with the line with the highest scrap rate

Reality: Vision AI doesn't replace the quality manager — it gives them X-ray vision. AI finds the defects, humans find the root causes.