Originally u-net neural network architecture was built for performing semantic segmentation on a small bio-medical data-set set [Ronneberger et al., 2015].
This deep neural networks is implemented with Keras.
Even though convolutional neural network (CNN) has recently become popular and has increasingly been used as an alternative to many traditional pattern recognition problems, it's application in material science is not common, for instance for segmenting stagnant zones of X-ray Computed Tomography (CT) images.
In the present work, a U-net architecture is used for segmenting the stagnant zone during silo discharging process. You can find the full conference paper : HERE.
As each of the CT images already contain repetitive structures with the corresponding variation, only very few images are required to train a network that generalizes reasonably well. As a result, to make the u-net architecture work with very few training images, it has been modified to provide more accurate segmentation.
Like on the original u-net, the modified architecture applied the same number of feature channels in upsampling part allow propagating context information to higher resolution layers.
One of the modifications is that the original u-net used stochastic gradient descent optimizer (Ronneberger et al., 2015), but this modified u-net architecture used Adam optimizer (Kingma and Ba, 2014) to minimize the categorical cross-entropy objective.
Since the available dataset is small, an extensive amount of data augmentations has been applied to improve the performance of the network.
The main goal of such augmentations is to prevent the network from memorizing just the training examples and to force it to learn about the stagnant zone boundaries. Therefore, in this study common transformation like rotation, flipping, shifting, zooming, and shearing are applied.
During training, the transformations are applied on the fly so that the network sees new random transformations during each epoch. Below table summarizes applied augmentation methods and you could apply/modify the augmentation in dataPrepare.ipynb.
| Method | Range |
|---|---|
| Rotation | +/-20 |
| Shift | 50% |
| Zoom | 50% |
| Shear | 50% |
The dataset was first divided into two subsets, train, and test. In which 80% were used for training and 20% for validation. The trained model was next tested on the second subset which contains 10 images.
The testing datasets were used for the evaluation of the network performance.
The model was trained for 5 epochs.
After 5 epochs, calculated accuracy (Intersection over Union (IoU)) was about 0.97 percent.
This tutorial depends on the following libraries:
-
Tensorflow
-
Keras >= 1.0
-
Python versions >=2.7
Put your training images in data/material/image Put your training labels in data/material/label
You could generate predicted results of test image in data/material/test
Use the trained model to generate predicted segmentation on test images. Some example images below: original and predicted segmentations:
Once after having trained model using the CNN method, the end-to-end 3D automatic segmentation offers an effective and fast segmentation of stagnant zone.
In order to prove that the trained model could generate the stagnant zone segmentation for completely different scan (the model was trained on sorghum grains flow scan), it was tested by using a 3D scan of rice grains flow. The result shows that the trained model was able to generate predicted segmentation successfully.
If you find our work useful, please cite:
@inproceedings{2019 IEEE Second International Conference on Artificial Intelligence and Knowledge Engineering (AIKE),
title = {Stagnant zone segmentation with U-net}
author = {Selam Waktola, Krzysztof Grudzien, Laurent Babout},
Conference = {Artificial Intelligence & Knowledge EngineeringAt: Cagliari, Italy},
year = {June 2019}
}
Kingma, D.P., Ba, J. 2014 : Adam: A Method for Stochastic Optimization. Int. Conf. Learn. Represent.