TrueDeep: A Systematic Approach of Crack Detection with Less Data

Installation

• Python >= 3.6 Recommended
• Install the requirements mentioned in the requirements.txt.

Datasets

• DeepCrack
  • This is a crack segmentation dataset that can be downloaded from the DeepCrack repository.
• kaggle-crack-segmentation
  • This is a crack segmetation dataset containing images from diverse backgrounds and is available here.
• FIVES
  • This is a vessel segmentation dataset that can be downloaded from here.
• BCCD
  • This is a blood cell segmetation dataset that can be downloaded from here.

Details

• We assume that a folder containing the dataset must be structured like {train-test-split}/{images-or-masks}/{filename.ext}, for example, train/images/10.jpg.

• subset folder contains the logic to obtain a coreset of images from an overall dataset. Choose any image dataset to subset your overall dataset and save the subset. Masks can also be saved if a binary segmentation dataset is used.

• segmentation folder contains the training methods to train the models for a binary segmentation task on the above mentioned datasets.

  • training folder contains the files required to perform training for binary segmentation, run train.sh with the required parameters to perform the training.
  • augmentation folder contains the implementation of augmentation techniques performed on crack masks, run augment_data.sh to perform the augmentation on the input images.
  • evaluation folder contains the implementation of inference and computation of metrics.
    • inference.py contains the code to perform inference using a trained model and inference can be performed using the script inference.sh.
  • F-score, Precision and Recall can be computed by running evaluate.sh. To compute more metrics like global accuracy and class average accuracy, please refer to the DeepSegmentor repository.

Model Checkpoints

• The model checkpoints are for truedeep are available here.

Workarounds

• If you're getting the error ImportError: cannot import name 'MultiHeadAttention' from 'tensorflow.keras.layers', comment out the corresponding lines of code from the keras_unet_collection source library.

Acknowledgements

We would like to thank the authors of the DeepCrack, FIVES and BCCD for their contributions to the research community and for making this research possible.

Reference

If you find the code useful for your research, please cite our paper:

@article{pandey2023truedeep,
  title={TrueDeep: A systematic approach of crack detection with less data},
  author={Pandey, Ramkrishna and Achara, Akshit},
  journal={Expert Systems with Applications},
  pages={122785},
  year={2023},
  publisher={Elsevier}
}

If you are using the stochastic width augmentations from the code, please cite our paper:

@inproceedings{pandey2023coredeep,
  title={CoreDeep: Improving crack detection algorithms using width stochasticity},
  author={Pandey, Ramkrishna and Achara, Akshit},
  booktitle={International Conference on Computer Vision and Image Processing},
  pages={62--73},
  year={2023},
  organization={Springer}
}