ML & DataInternship · HAMK Smart2024Completed
Blister Detection on Coated Metal
Industrial quality inspection with computer vision (HAMK Smart internship).
- precision (YOLOv8s, synthetic-data training)
- 97.5%
- precision (YOLOv8s, synthetic-data training)
- mAP@50
- 90.5%
- mAP@50
Overview
Detecting and classifying blisters on colour-coated metal plates. During the internship I compared classical machine vision (OpenCV) with deep-learning detection, and built a Django web app for comparing samples with standard reference images. A follow-up YOLOv8 study combined synthetic and real training data.
Notes
- The reported metrics are from the model trained on synthetic data. A separate model was fine-tuned for real-world images.
- Recall (85.3%) is lower than precision, so the model misses some small blisters.