The goal of this project is to develop an image segmentation model for detecting cracks in images. Unlike fully supervised approaches that require pixel-level annotations for training, this project explores a weakly supervised segmentation setup, where only image-level labels are available during training.
Specifically, the model is trained using images labeled as cracked or not cracked, without providing explicit crack masks during the training process. Despite this limitation, the model learns to localize crack regions by leveraging architectural inductive biases and weak supervision techniques.
The segmentation model is based on a Weakly Supervised U-Net architecture with a ResNet-50 encoder.
The model was trained and evaluated using the Crack Segmentation Dataset available on Kaggle:
- Dataset: Crack Segmentation Dataset
- Source: https://www.kaggle.com/datasets/lakshaymiddha/crack-segmentation-dataset
- Training images: 9,603
- Testing images: 1,695
- Training epochs: 3
- Supervision: Image-level labels only (
cracked/not cracked) - Evaluation metric: Dice Coefficient (Dice Similarity Coefficient, DSC)
The final performance on the test set achieved a Dice score of 0.3908.
The Dice score measures the overlap between predicted segmentation masks and ground-truth masks and is defined as:
-
Ronneberger, O., Fischer, P., & Brox, T. (2015).
U-Net: Convolutional Networks for Biomedical Image Segmentation.
arXiv preprint arXiv:1505.04597.
https://arxiv.org/abs/1505.04597 -
Yuan, Y., et al. (2022).
Weakly Supervised Learning for Crack Segmentation.
bioRxiv.
https://www.biorxiv.org/content/10.1101/2022.09.09.507144v1 -
Towards Data Science.
Cook Your First U-Net in PyTorch.
https://towardsdatascience.com/cook-your-first-u-net-in-pytorch-b3297a844cf3/ -
Turgutlu, K.
Weakly Supervised Transfer Learning in Medical Imaging.
https://medium.com/@keremturgutlu/weakly-supervised-transfer-learning-in-medical-imaging-c89c5ca2d0be