Academic project · BUas Academic
Wall Plug Defect Detection
Built a computer-vision defect-detection system for manufacturing quality control that hit 100% test accuracy across four wall-plug types — matching commercial systems costing $15,000+ while running on a Raspberry Pi at roughly 10% of the cost.
Context
Manual visual inspection on a production line is slow, monotonous and error-prone — fatigue alone lets defects through, driving waste and rework. As an academic project at BUas, I set out to reconstruct industrial-grade defect detection cheaply enough to be realistic for small and medium manufacturers, and transparent enough to trust on a factory floor.
Approach
I worked against a human-level accuracy benchmark of 95% and improved the model in stages:
- Baseline — a simple feedforward network established the floor at 84% accuracy.
- Custom CNN — purpose-built for wall-plug features, reaching 90%.
- Transfer learning — data augmentation plus pre-trained MobileNetV2 and EfficientNetB0 (both edge-friendly) pushed the model to 100% accuracy on the test set across four plug types: frame fixing, insulation anchors, ribbed plugs and toggle anchors.
Trust & deployment
A black-box pass/fail isn’t enough on a production line, so I integrated Grad-CAM heatmaps to show which regions drove each decision, and wrapped everything in a desktop app with single-image and batch classification, confidence scores and decision logs for audit. User testing with potential operators then guided the interface refinements.
Results
- 100% test accuracy, exceeding the 95% human benchmark.
- Real-time inference on a Raspberry Pi 4, at roughly 10% of commercial-system cost (~$1,500 vs $15,000+).
What I took from it
This project convinced me that explainability isn’t a nice-to-have in industrial ML — the Grad-CAM views were what made the system trustworthy — and that modern architectures can run genuinely useful vision on low-cost edge hardware.