Semiconductor Defect Detection
Key Results
- 98% precision in defect identification
- 0.2-second inspection time per image
- Real-time factory floor deployment
- Complete ML workflow with WPF application
The Problem
Semiconductor manufacturing requires extremely precise quality control. Manual visual inspection was slow, inconsistent, and unable to keep pace with production line speeds. Even small defects missed during inspection could result in costly downstream failures.
The Approach
Developed a real-time defect detection system using YOLOv3 and RetinaNet models trained on annotated semiconductor wafer images. The pipeline included automated image capture from factory floor cameras, preprocessing with OpenCV, and inference optimised for sub-second response times. A C#/WPF desktop application provided the operator interface for real-time monitoring and defect logging.
The Results
The system achieved 98% precision in defect identification with 0.2-second inspection time per image, enabling real-time deployment on the factory floor. The complete ML workflow with integrated WPF application replaced manual inspection, dramatically improving throughput and consistency.