This repository contains the codes for crack detection using topological loss function. The methodology hereby implemented was presented in the paper "TOPO-Loss for continuity-preserving crack detection using deep learning" by Pantoja-Rosero et., al. (2022)
Clone repository in your local machine. All codes related with method are inside the src directory.
Example input data can be downloaded from Dataset for TOPO-Loss for continuity-preserving crack detection using deep learning. This datased contains 3 main folders. data, models results. Extract the three folders and place them inside the repository folder
The repository directory should look as:
topo_crack_detection
└───data
└───docs
└───models
└───results
└───src
Create a conda environment and install python packages. At the terminal in the repository location.
conda create -n topo_crack_detection python=3.7
conda activate topo_crack_detection
pip install -r requirements.txt
pip3 install torch torchvision
Open the terminal inside the src folder (with the environment activated -- conda activate topo_crack_detection) and write the next command:
python test.py
The script by default will call the MSE+TOPO trained model and used it with the full sized images placed inside data\test_set\images\. The inference results will be placed inside the folder results\.
Note: If want to test another model, change the path inside the test.py file. If your memory is overflowed, reduce the patch size
To train the models with the provided data set, ppen the terminal inside the src folder (with the environment activated -- conda activate topo_crack_detection) and write the next command:
python main.py --model_name="your_modeld_choice" --lr=your_learning_rate --n_epoch=your_epoch_number --malis_neg=your_topo_parameter1 --malis_pos=your_topo_parameter1
The next lines are used to train some of the models presented in the paper.
- MSE model
python main.py --model_name="mse" --lr=5e-6 --n_epoch=50 - TOPO model
python main.py --model_name="topo" --lr=3e-5 --n_epoch=50 --malis_neg=100 --malis_pos=10 - DICE+TOPO model
python main.py --model_name="dice+topo" --lr=3e-5 --n_epoch=50 --malis_neg=100 --malis_pos=10 - MSE+TOPO model
python main.py --model_name="mse+topo" --lr=3e-5 --n_epoch=50 --malis_neg=100 --malis_pos=10
Note: malis_neg -> stimulates connectivity. malis_pos -> helps to decrease false-positives. The saved models will be placed inside the models/ folder.
Follow the structure of the data/ folder and place your data accordingly (training and validation datasets). Note that the ground truth used is derivated from the skeleton of the crack annotation. Train the models following the commands described in 5.
Once you have trained your modes, change the model's path accordingly inside the file test.py. Run the command as described in 4.
The results will be saved inside results folder. This are formed by predictions as distance maps, thresholded binary image and original images with damage overlayed.
We kindly ask you to cite us if you use this project, dataset or article as reference.
Paper:
@article{Pantoja-Rosero2020a,
title = {TOPO-Loss for continuity-preserving crack detection using deep learning},
journal = {Construction and Building Materials},
volume = {344},
pages = {128264},
year = {2022},
issn = {0950-0618},
doi = {https://doi.org/10.1016/j.conbuildmat.2022.128264},
url = {},
author = {B.G. Pantoja-Rosero and D. Oner and M. Kozinski and R. Achanta and P. Fua and F. Perez-Cruz and K. Beyer},
}
Dataset:
@dataset{Pantoja-Rosero2022a-ds,
author = {Pantoja-Rosero, Bryan German and
Oner, Doruk and
Kozinski, Mateusz and
Achanta, Radhakrishna and
Fua, Pascal and
Perez-Cruz, Fernando and
Beyer, Katrin},
title = {{Dataset for TOPO-Loss for continuity-preserving
crack detection using deep learning}},
month = jun,
year = 2022,
publisher = {Zenodo},
version = {v0.0},
doi = {10.5281/zenodo.6769028},
url = {https://doi.org/10.5281/zenodo.6769028}
}