# Reducing Scrap Rates by 60%: Computer Vision QC Implementation
Scrap is manufacturing's silent profit killer. As Deloitte's smart factory research indicates, computer vision-powered quality control is transforming how manufacturers detect defects—earlier, more consistently, and more accurately than human inspection.
The Scrap Problem Quantified
True Cost of Scrap
| Cost Category | Typical Impact |
|---|---|
| Material | 30-40% of scrap cost |
| Labor | 20-30% |
| Machine time | 15-25% |
| Energy | 5-10% |
| Overhead | 10-20% |
Why Traditional QC Falls Short
Human Inspector Limitations - Fatigue after 20-30 minutes of concentrated inspection - Subjective judgment varies between inspectors - Limited to visible defects, one angle at a time - Difficult to scale during demand spikes
> Download our free Industry 4.0 Readiness Assessment — a practical resource built from real implementation experience. Get it here.
## Computer Vision QC Fundamentals
How It Works
- 1Image Capture: High-resolution cameras photograph each part
- 2Preprocessing: Images normalized, enhanced, aligned
- 3AI Analysis: Deep learning models detect anomalies
- 4Decision: Pass, fail, or hold for review
- 5Action: Reject mechanism, alert, or data logging
Defect Types Computer Vision Excels At
Surface Defects: Scratches, dents, discoloration, contamination Dimensional Issues: Missing features, size tolerance, warpage Assembly Verification: Missing components, wrong orientation Label and Print: Missing labels, print quality, wrong barcodes
Technology Components
Cameras - Area scan: Standard 2D imaging - Line scan: For continuous materials - 3D structured light: Dimensional and surface - Hyperspectral: Material composition
Lighting - Diffuse: Minimizes reflections - Directional: Highlights scratches - Backlighting: Silhouette analysis
Implementation Roadmap
Phase 1: Assessment (4-6 weeks)
Current State Analysis - Document current scrap rates by product and defect type - Analyze inspection costs and limitations - Map production flow and inspection points
Use Case Prioritization
| Criteria | Weight |
|---|---|
| Scrap cost | 25% |
| Detection difficulty | 20% |
| Implementation feasibility | 20% |
| Quality impact | 20% |
| Production impact | 15% |
Phase 2: Pilot Design (2-4 weeks)
Image Collection Requirements
| Category | Images Needed | Purpose |
|---|---|---|
| Good parts | 500-2000 | Define normal |
| Defective parts by type | 100-500 each | Train detection |
| Borderline cases | 50-200 | Define threshold |
Hardware Selection
| Requirement | Camera Type | Resolution |
|---|---|---|
| Fast line, simple defects | Area scan | 2-5 MP |
| Web inspection | Line scan | 4-8K pixels |
| 3D dimensional | Structured light | 1-2 MP |
Phase 3: Development (6-10 weeks)
Model Development Options
| Option | Pros | Cons |
|---|---|---|
| Pre-built Platforms (Cognex, Keyence) | Faster, proven | Limited customization |
| Custom Development (TensorFlow) | Maximum flexibility | Requires ML expertise |
| Hybrid Approach | Best of both | More complex |
Training Strategy 1. Initial training with collected images 2. Active learning with edge case review 3. Continuous improvement from production data
Phase 4: Deployment (4-6 weeks)
Shadow Mode Operation - Run parallel to existing inspection - Compare AI decisions to human decisions - Build confidence before going live
Key Metrics - True positive rate - False positive rate - True negative rate - False negative rate (critical!)
Recommended Reading
- Automotive Supplier Reduces Defects by 73% with AI Quality Inspection: A Manufacturing Success Story
- Computer Vision Quality Control: Building Defect Detection Systems with 99.8% Accuracy
- Connecting Legacy PLCs to AI Systems: OT/IT Integration Guide
## Case Studies
Automotive Stamping - 3 stamping lines, 4.2% scrap rate - 12-camera system with multi-angle lighting - **Result**: Scrap reduced to 1.1% (74% reduction), $1.7M annual savings
Electronics PCB Assembly - SMT assembly line, 2.8% defect rate - AI-enhanced AOI with 3D solder inspection - **Result**: Final test defect rate 0.4% (86% reduction)
Food Packaging - 200 packages/minute high-speed line - Line scan cameras for label and seal inspection - **Result**: Retail rejection rate 0.2% from 2.1%
Common Challenges
Challenge 1: High False Positive Rates
Causes and Solutions - Lighting inconsistency → Improve enclosure - Normal variation not in training → Expand training data - Threshold too aggressive → Adjust confidence threshold
Challenge 2: Missed Defects
Causes and Solutions - Defect type not in training → Add examples, retrain - Camera resolution insufficient → Upgrade cameras - Lighting doesn't reveal defect → Test alternative lighting
Challenge 3: Production Speed Mismatch
Causes and Solutions - Processing too slow → Upgrade edge hardware - Image transfer bottleneck → Upgrade interface - Model too complex → Simplify architecture
Technology Selection Guide
Vendor Comparison
Turnkey Solutions - Cognex: Comprehensive platform, premium pricing - Keyence: Excellent support, good ease-of-use - SICK: Strong industrial integration
Platform Providers - MVTec HALCON: Industry standard, flexible - Open source (OpenCV + TensorFlow): Maximum flexibility
ROI Calculator
Investment
| Item | Typical Range |
|---|---|
| Camera system per station | $15,000-$50,000 |
| Lighting | $2,000-$15,000 |
| Edge processing | $5,000-$25,000 |
| Software licenses | $10,000-$50,000 |
| Integration labor | $20,000-$100,000 |
Value Calculation ``` Annual Savings = Current Scrap Rate × Reduction % × Volume × Unit Cost
Example: - Scrap rate: 3%, Reduction: 60%, Volume: 1M units, Unit cost: $25 - Annual savings: 0.03 × 0.60 × 1,000,000 × $25 = $450,000 ```
Getting Started
- 1Measure current state: What's your scrap rate by defect type?
- 2Prioritize opportunities: Which defects have highest impact?
- 3Collect images: Start building your training dataset
- 4Evaluate technologies: Demo solutions on your actual parts
- 5Plan pilot: Scope a contained implementation
Contact APPIT's manufacturing AI team for a computer vision QC assessment.
