Multi-scale Convolutional Neural Networks (MCNN)

---

MCNN extends the functionality of the hidden layers in the decoder of a U-Net by connecting them to additional convolution layers to produce coarse outputs, in attempt to match the low-frequency components. This greatly accelerates the convergence and enhances the stability of the neural-network. The convergence curve with U-net is shown in the figure blow.

Webpage

System Requirement (Training)

• software (python packages)

  • Python (3.7.4)
  • Tensorflow (1.14.0)
  • Keras (2.2.4)
  • numpy (1.17.0)
  • openCV (4.1.1)
  • scikit-image (0.15.0)
  • tifffile (0.9.2)
  • imageio (2.0.10)
  • jupyter-notebook (6.0.1)
  • mss (4.0.0)
  • scipy (1.3.1)
  • matplotlib (3.1.1)

• recommanded hardware

  • 2T hard disk space (for simulated dataset)
  • 256 GB Memory
  • 2xGTX 1080 Ti GPU

Demo/Quick Tutorial

For people who is interested in applying MCNN to their own project, check out the tutorial folder.

Phase retrieval applications

For the phase retrieval applications, please check out folder phase_retrieval;

Imageing objects from diffuse reflection

For the imaging objects from diffusive reflection application, please check out folder diffuse_reconstruction;

Denoising STEM images

For the STEM images denoising application, please check out folder denoising;

Pre-trained models

Some of the pre-trained models can be found in MCNN-DEMO repo.

bibtex

We kindly ask you to cite our publication if you use our dataset, code or models in your work.

@article{wangMultiresolutionConvolutionalNeural2020,
  title = {Multi-Resolution Convolutional Neural Networks for Inverse Problems},
  author = {Wang, Feng and Eljarrat, Alberto and Müller, Johannes and Henninen, Trond R. and Erni, Rolf and Koch, Christoph T.},
  date = {2020-03-31},
  journaltitle = {Scientific Reports},
  volume = {10},
  pages = {1--11},
  issn = {2045-2322},
  doi = {10.1038/s41598-020-62484-z},
  url = {https://www.nature.com/articles/s41598-020-62484-z},
  urldate = {2020-04-01},
  langid = {english},
  number = {1}
}

License

AGPLv3