Concrete Crack Segmentation using U-Net

This repository provides an implementation of an Improved U-Net architecture for concrete crack segmentation.
The model is optimized for highly imbalanced datasets, where cracks make up a very small portion of the image compared to the background.

• Improved U-Net with dropout & batch normalization
• Multiple loss functions for imbalanced data: Dice, Focal, Tversky, Weighted BCE, and Combined loss
• Training pipeline with class balancing, learning rate scheduling, gradient clipping, and early stopping
• Support for BCEWithLogitsLoss (logits output variant of U-Net included)
• Designed for detecting small cracks in large concrete surfaces

Deployed on Hugging Face at

https://huggingface.co/spaces/faranbutt789/Unet_crack_detection