Visual inspection for surface defects

The brief

A manufacturer inspects parts by eye at the end of the line, and defects still slip through. Build a vision system that learns from mostly good parts, highlights where a defect is on the image, and flags parts for rejection. Good looks like: strong defect detection with the defect highlighted, a rejection threshold chosen with the production team in mind, and a working inspection demo.

What you will build

1. Explore the image categories and defect types
2. Train an anomaly-detection model on good parts only
3. Produce heat maps showing where the defect is
4. Choose a threshold from the cost of missed defects vs false rejects
5. Use tabular fault data to compare with a classical approach
6. Build an inspection demo and write the report

Free, open data to build it with

• MVTec Anomaly Detection (MVTec AD) ↗ MVTec
• Steel Plates Faults ↗ UCI Machine Learning Repository

Check each dataset's licence before you publish your work. Your mentor gives your team its own version of the brief (data, scope or region), so every team's project is its own.