Project Overview
We built a computer-vision system that inspects products on the line in real time, detecting defects and classifying items with accuracy and consistency that manual inspection could not match.
The Challenge
Manual visual inspection was slow, fatiguing, and inconsistent. Defects slipped through to customers, and there was no data on where quality issues originated.
- Manual inspection was slow and inconsistent
- Defects escaped to customers, hurting reputation
- Inspector fatigue caused missed detections
- No analytics on defect types or sources
Our Strategic Approach
We trained custom detection and classification models on labeled defect imagery and deployed them at the edge for real-time, low-latency inspection on the line.
The Solution We Delivered
The system flags defects instantly, classifies type and severity, and feeds a dashboard that pinpoints where and why quality issues arise.
- Real-time defect detection and classification
- Edge deployment for low-latency inspection
- Severity scoring and automatic reject signals
- Defect analytics by type, line, and shift
- Active-learning loop from flagged edge cases
- Integration with line-control systems
Technologies Used
- PyTorch — Detection and classification model training
- YOLO / CNNs — Real-time object and defect detection
- ONNX / TensorRT — Optimized edge inference
- Python — Vision pipeline and tooling
- React — Inspection analytics dashboard
- Edge devices — On-line, low-latency inference
Development Process
- Data capture & labeling — Collected and annotated defect imagery across conditions.
- Model training — Trained and validated detection and classification models.
- Edge optimization — Quantized and optimized models for real-time inference.
- Line integration — Connected to cameras and reject mechanisms.
- Active learning — Looped flagged edge cases back into training.
Results & Impact
Inspection became fast, consistent, and data-rich, catching defects manual review missed.
- Defect detection accuracy above 98%
- Inspection throughput increased significantly
- Customer-reported defects reduced sharply
- Actionable analytics on defect root causes
🎯 Key Takeaway
Computer-vision inspection delivered consistent, real-time quality control and turned visual data into insight that drives upstream improvements.
Project Gallery
