This repository is the official implementation of the Crack Segmentation U-shape (CrackSeU) Network.
Our paper is finally accepted by Automation in Construction after a year of review. I have to say it has been a long and tough journey. 😭.
The paper is available:
Online monitoring of crack dynamic development using attention-based deep networks, Automation in Construction, 154 (2023) 105022, by Wang chen*, Zhili He*, and Jian Zhang#. ( *: Co-first authors, #: Corresponding Author )
Recommended versions are
* python = 3.5
* pytorch = 1.12.1
* CUDA 11.6.2 and CUDNN 8.6.0
Other requirements can be found in the Requirements.txt.
git clone https://github.com/hzlbbfrog/CrackSeU
cd CrackSeU
pip install -r Requirements.txtOr, you can directly "Download ZIP".
You can refer to the following file tree to organize your own data.
Your project
│ README.md
│ ...
│ CrackSeU_main.py
│
└───Dataset
|
└───Your dataset name
|
└───Train
└───images
└───masks
└───Test
└───images
└───masks
│
└───...Other directories
- Include CrackSeU-B with LN_VT.
- Include CrackSeU-B with BN.
- Include CrackSeU-B with LN_Pytorch.
- Include CrackSeU-B with LN_He.
To train the CrackSeU-B with LN_VT, simply run:
python CrackSeU_main.py --action=train --arch=CrackSeU_B_LN_VT --epoch=50 --batch_size=2 --lr=1e-4To test the CrackSeU-B with LN_VT, simply run:
python CrackSeU_main.py --action=test --arch=CrackSeU_B_LN_VT --test_epoch=50
| Method | m IoU (%) | mi IoU (%) | mi Dice (%) | #Param. (M) | MACs (G) |
|---|---|---|---|---|---|
| U-Net | 81.04 | 75.35 | 81.20 | 7.77 | 55.01 |
| U-Net (large) | 82.65 | 76.18 | 81.40 | 31.04 | 219.01 |
| U-Net++ | 79.51 | 74.02 | 80.14 | 9.16 | 138.63 |
| U-Net++ (large) | 80.33 | 74.50 | 81.03 | 36.63 | 552.67 |
| Attention U-Net | 82.87 | 75.85 | 81.17 | 34.88 | 266.54 |
| CE-Net | 81.28 | 75.25 | 81.09 | 29.00 | 35.60 |
| CrackSeU-B | 85.74 | 81.32 | 88.55 | 3.19 | 11.22 |
| CrackSeU-M | 85.85 | 81.53 | 88.66 | 3.58 | 15.04 |
| CrackSeU-L | 86.39 | 82.09 | 89.11 | 4.62 | 28.22 |
It is worth noting that the number of parameters of CrackSeU-L is 4.62M.
In the original paper, we mistakenly considered the parameters of the SOB so that the data is 4.70M and a little higher than the true #Param. (4.62M).
We are really sorry if this makes you confused.
| Method | m IoU (%) | mi IoU (%) | mi Dice (%) | F1 score | #Param. (M) | MACs (G) |
|---|---|---|---|---|---|---|
| U-Net | 69.41 | 68.17 | 75.07 | 78.16 | 7.77 | 43.84 |
| U-Net (large) | 69.61 | 68.40 | 75.64 | 78.41 | 31.04 | 174.53 |
| U-Net++ | 70.19 | 67.92 | 74.91 | 78.20 | 9.16 | 110.47 |
| Attention U-Net | 71.48 | 69.19 | 75.11 | 79.16 | 34.88 | 212.40 |
| CE-Net | 69.24 | 68.80 | 76.10 | 79.30 | 29.00 | 28.37 |
| DeepLabv3+ (MobileNetv2) | 69.70 | 69.18 | 74.23 | 78.34 | 5.81 | 23.25 |
| DeepLabv3+ (ResNet-101) | 70.15 | 67.52 | 73.82 | 78.38 | 59.34 | 70.80 |
| CrackSeU-B | 73.80 | 71.71 | 81.40 | 81.82 | 3.19 | 8.94 |
Correction on 2024/12/24
We find a difference between the above results and the table in the original paper. This mistake is because when we do the table of Deepcrack, we first copy the table of Concretecrack and then change the values. We missed changing the value (belonging to Concretecrack) to the new one (Deepcrack). We have corrected it.
You are very welcome to cite our paper! The BibTeX entry is as follows:
@article{CrackSeU,
title = {Online monitoring of crack dynamic development using attention-based deep networks},
journal = {Automation in Construction},
volume = {154},
pages = {105022},
year = {2023},
doi = {https://doi.org/10.1016/j.autcon.2023.105022},
url = {https://www.sciencedirect.com/science/article/pii/S0926580523002820},
author = {Wang Chen and Zhili He and Jian Zhang},
keywords = {Crack identification, Online monitoring method, Deep learning}
}SEU is also the abbreviation of Southesast Univertisy.
The name of our framework ( CrackSeU) is also dedicated to the 120th anniversary of Southeast University.