AI/ML & Data
Visual Defect Detection
A CNN that spots manufacturing defects the eye misses at speed.
A convolutional model that classifies product images as defective or clean — consistent inspection that never gets tired at piece 4,000.
01
The problem
Manual visual inspection degrades with fatigue and varies by inspector — defects pass, and good parts get rejected.
02
Our approach
A CNN trained on labeled defect imagery with augmentation for robustness, evaluated per defect class rather than on one blended score.
03
The outcome
Reliable defect classification suitable as a first-pass screen ahead of human review.
Highlights
- Per-class evaluation
- Augmentation for robustness
- First-pass screening design
