FCN_for_crack_recognition

Requirements

• Python 3.x
• Tensorflow >= 1.21 or Tensorflow-gpu
• Numpy
• Scipy, Scikit-image
• Matplotlib

Content

• FCN_DatasetReader.py: Classes for training dataset and test image reading
• FCN_layers.py: Functions of layers
• FCN_model.py: Model of FCN
• FCN_finetune.py: Main training and test of FCN
• data/train/*: Folder for training dataset, contains subfolder 'image', 'annotation' and 'index.txt'
• data/valid/*: Folder for validing dataset, contains subfolder 'image', 'annotation' and 'index.txt'
• logs: Folder for training logs
• checkpoints: Folder for model parameters
• test: Folder for test images

Useage

Test

1. Download pretrained model (https://drive.google.com/open?id=1oX7IO0R_ZkfHwZ_zV4c3_v9_24empwNz) and put into folder checkpoints
2. Put test images into folder test
3. Run python FCN_finetune.py --mode=predict --test_dir=test

Train and finetune

1. Download vgg19 pretrained parameters into the root folder (https://drive.google.com/open?id=15WMDJbFWw3f1qMbTuDO1k4HQ0hyPB4-6)
2. Prepare your own data or download crack dataset from (https://drive.google.com/open?id=1cplcUBmgHfD82YQTWnn1dssK2Z_xRpjx) If you need to change the training samples or validating sample, you can modify the index.txt file directly. Then put the data into data/train/ and data/valid/ respectively.
3. Run python FCN_finetune.py --mode=finetune --learning_rate=1e-4 --num_of_epoch=20 --batch_size=2
4. If you would like to check the training process, run tensorboard --logdir=logs, then open http://localhost:6006/ using any web explorer.

Please put 'index.txt' into train or valid folder as follows (The feeding process will follow this order):

image//0002.jpg,annotation//0002.png
image//0001.jpg,annotation//0001.png

Skeleton of cracks

Once you have got the predictions of cracks, go to python environment

from FCN_CrackAnalysis import CrackAnalyse

analyser = CrackAnalyse('test/001.png')
crack_skeleton = analyser.get_skeleton()
crack_lenth = analyser.get_crack_length()
crack_max_width = analyser.get_crack_max_width()
crack_mean_width = analyser.get_crack_mean_width()

Then you can using matplotlib to show the skeleton and print the crack morphological features.

Results

• training loss

• training accuracy

• normal cracks

• thin cracks

• intersected cracks

• historical(wide) cracks

• mixed cracks

• complex cracks