Segmentation Architectures

This repository contains multiple architectures for segmentation. The architectures are:

• UNet (from https://arxiv.org/abs/1505.04597) with a lot of additions, which include
  • Residual connections
  • Attention
  • Squeeze and Excitation blocks
  • CBAM blocks
• Deeplabv3+ (from https://arxiv.org/abs/1802.02611v2)
• DeepTiramisu (from https://arxiv.org/abs/1611.09326)

If you use our code in your work please cite the following papers:

• Albert, S.; Wichtmann, B.D.; Zhao, W.; Maurer, A.; Hesser, J.; Attenberger, U.I.; Schad, L.R.; Zöllner, F.G. Comparison of Image Normalization Methods for Multi-Site Deep Learning. Appl. Sci. 2023, 13, 8923. https://doi.org/10.3390/app13158923

Getting Started

The models can be build by calling the build_model function in the submodule corresponding to the model.

Prerequisites

No prerequisites are required besides the modules listed in the requirements.txt file.

Installing

It is best to use virtualenv to create a virtual environment

python -m virtualenv venv

And then install the requirements.

Pre-commit can be installed with

pip install pre-commit

The cooks will be installed by

pre-commit install

You can run the hooks for all files using (usually, they are run only for files being committed)

pre-commit run --all-files

Running the tests

• The test can be run using pytest
• They can also be run by hand using python -m SegmentationArchitectures.test_architectures

Running the training

For training, just compile the model and train it with model.fit or a custom training function (easiest way is to subclass the tf model)