Defect Detection with Deep Learning Computer Vision in Jetson Nano
Real-time defect detection system for Quality Control using Jetson Nano
Real-Time Defect Detection on Jetson Nano
Developed a YOLOv4-based deep learning system running at 29.8 FPS on edge hardware for automated quality control inspection, tested as a working prototype at the QC station.
The Problem
At PT Toshin Prima Fine Blanking, quality control was done manually by human operators — slow, inconsistent, and prone to fatigue-related errors. The factory needed an automated, real-time solution that could run on affordable hardware without cloud dependency.
What I Built
An end-to-end visual inspection pipeline: from data collection and model training to prototype testing.
- Collected and labeled product defect datasets from the actual production line
- Trained a custom YOLOv4 model optimized with TensorRT for edge inference
- Integrated with Logitech Brio camera for high-resolution image capture
- Built a Python application for real-time visualization, classification, and logging
- Tested the system on NVIDIA Jetson Nano at the QC station as a working prototype
YOLOv4 TensorRT Jetson Nano Python OpenCV Deep Learning Edge AI
Results
| Metric | Value |
|---|---|
| Inference Speed | 29.8 FPS |
| F1-Score | 0.9945 |
| Deployment | Prototype (QC station) |
| Hardware Cost | ~$150 (Jetson Nano) |
- Detected multiple defect types with near-perfect accuracy
- Reduced manual inspection time significantly
- System ran fully offline — no cloud required
Detection Example
