https://github.com/saroshfarhan/architectures-for-semantic-segmentation-using-deep-learning
Implementation of various semantic segmentation architectures in Keras using Tensorflow backend
https://github.com/saroshfarhan/architectures-for-semantic-segmentation-using-deep-learning
Last synced: 19 days ago
JSON representation
Implementation of various semantic segmentation architectures in Keras using Tensorflow backend
- Host: GitHub
- URL: https://github.com/saroshfarhan/architectures-for-semantic-segmentation-using-deep-learning
- Owner: saroshfarhan
- Created: 2018-01-23T13:42:34.000Z (over 8 years ago)
- Default Branch: master
- Last Pushed: 2018-01-23T08:20:36.000Z (over 8 years ago)
- Last Synced: 2025-03-05T11:20:08.846Z (over 1 year ago)
- Language: Python
- Homepage:
- Size: 14.1 MB
- Stars: 0
- Watchers: 3
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Architectures-for-Semantic-Segmentation-using-Deep-Learning
This repository contains the implementations of [U-Net](https://lmb.informatik.uni-freiburg.de/people/ronneber/u-net/) and [V-Net](https://arxiv.org/abs/1606.04797) in Keras using TensorFlow backend. Also, `data_prep.py` contains the code for random elastic deformations applied to the input images for data augmentation, which were specified to be of importance in both the papers.
## Data
The `data` folder contains pre-processed images, converted to .tif format from 3-D volume tiff as present in the [challenge website](http://brainiac2.mit.edu/isbi_challenge/).
## Architectures
- ### U-Net
**Highlights**
- Encoder-decoder architecture as proposed in [Fully Convolutional Networks for Semantic Segmentation
](https://arxiv.org/abs/1605.06211)
- Each stage in the encoder part has 2 3x3 conv layers followed by a MaxPooling layer, which performs the downsampling operation.
- Number of filters double in each subsequent stage of the encoder part.
- In the decoder part, transpose convolution is used for upsampling and high-level features of the corresponding stage in the encoder part are concatenated at the beginning of each stage.
- The last layer uses 1x1 convolutions to give the final segmentation map.
- Extensive use of data augmentation, specially random elastic transformations.
- ### V-Net
**Highlights**
- Similar architecture to U-Net, but with residual connections (https://arxiv.org/abs/1512.03385) between the input and output of each stage.
- Downsampling performed by strided convolution instead of MaxPooling.
- Varying number of convolutional layers in each stage.
- PReLU used as the activation.
Although the original papers don't specify, I have always found **Batch Normalization** to speed up the training process and hence, the implementation includes that. Also, **Dropout** is applied between the convolutional layers.
## Running the model
First, prepare the augmented images by running:
```bash
python data_prep.py
```
Then, run the script corresponding to the architecture. E.g. for training U-Net, run the following:
```bash
python run_unet.py
```
## Dependencies
- TensorFlow
- Keras
- Numpy
- OpenCV
Install them using [pip](https://pypi.python.org/pypi/pip).
## Contributing
Feel free to create a Pull Request to add other semantic segmentation architectures and/or benchmark results or various datasets. If you are a beginner, you can refer [this](https://opensource.guide/how-to-contribute/) for getting started.
## Support
If you found this useful, please consider starring(★) the repo so that it can reach a broader audience.
## References
The pre-processed images and most part of `data_prep.py` were taken from [here](https://github.com/zhixuhao/unet).