Virginia Tech
2 files

Trained model for the Semantic Segmentation of Concrete Cracks (LCW)

posted on 2021-10-06, 13:17 authored by Eric Bianchi, Matthew Hebdon

This contains four trained DeepLabV3+ models for the semantic segmentation of concrete cracks. The models all were trained using image sizes of 512x512. The models were trained for 20 epochs each and the best epoch using the validation data during training was kept. The classes for this model are: [background, and crack]. This model was trained using the Labeled Cracks in the Wild (LCW) dataset (DOI: 10.7294/16624672) and with the conglomerate concrete crack dataset (DOI: 10.7294/16625056). The four models include: [LCW_cracked (no blank images), LCW_pretrained_crack (pretrained weights of conglomerate and no blank LCW images), LCW_pretrained (pretrained weights of conglomerate and LCW images), and LCW (LCW images)]. The best model, LCW_cracked, was able to correctly identify approximately 40% of the ground truth labeled cracks in the test set. More details of the training, the results, the dataset, and the code may be referenced in the journal article. The repository of the training and testing code may be accessed using the GitHub repository referenced in the journal article.

If you are using the model in your work, please include both the journal article and the model citation.


National Science Foundation Grant No. IIS-1840044



University Libraries, Virginia Tech


  • English (US)


Virginia Tech