Deep Learning Shipped · 2024
Concrete Crack Segmentation
U-Net semantic segmentation for infrastructure inspection
Local project
Deep Learning 2024
System architecture
- 01
Dataset
Concrete-surface imagery
- 02
Model
U-Net implemented and trained from scratch (PyTorch)
- 03
Segmentation
Pixel-level crack masks
- 04
Evaluation
Precision-oriented metrics — IoU, Dice
Pixel-level segmentation of cracks on concrete surfaces using U-Net. Trained from scratch, with particular care for precision: in infrastructure inspection, a false negative has a real cost.
What was built
- U-Net architecture implemented and trained from scratch
- Concrete-surface image dataset
- Precision-oriented segmentation metrics (IoU, Dice)
- Use case: automated infrastructure inspection
The real challenge
U-Net is still a reference for segmentation: understanding why skip connections preserve spatial detail taught me as much as the training itself.
Next project
MarketPulse AI