We highly recommend using the Linux operating system. Using an nVIDIA GPU with >=11GB memory is highly recommended, although this tool can be used with CPU only.

We used Ubuntu 9.4.0 (GNU/Linux 5.15.0-76-generic x86_64)) and an NVIDIA Corporation Device 2204 (rev a1) GPU and CUDA version 11.4.

• Python 3.7

• Packages:

  basicsr == 1.4.2

  easydict == 1.11.dev0

  imageio == 2.13.3

  keras == 2.11.0

  numpy == 1.21.5

  opencv-python == 4.5.4.60

  Pillow == 9.0.1

  scikit-image == 0.19.2

  scipy == 1.7.3

  tensorflow-gpu == 2.7.0

  tifffile == 2021.11.2

  torch == 1.10.0+cu113

  csbdeep

The overall training process can be divided into two phases. In the first training phase, the model is optimized with an uncertainty-driven loss (UDL) to learn the uncertainty of the IR results and the common L2 norm loss to calculate the similarity between the IR results and the HR images. In the second training phase, the uncertainty-weighted (UWL) loss fine-tunes the model to focus more on image areas with larger reconstruction uncertainty. The well-trained model in the first phase is frozen to provide the uncertainty for calculating the weight in UWL loss.

• The first stage of training the model for denoising

cd ./denoise/
python training_TF.py

• Modify training_TF.py istrain=True modeltypelst = ['FBuncertainty_f_uncer'] Replacing "traindatapath" with the path of the training data data_label.npz.

• The second stage of training the model for denoising

cd ./denoise/
python training_TF.py

• Modify training_TF.py modeltypelst = ['FBuncertainty_f_twostage'] Replacing "modelpath1" with the path of the model trained in the first stage.

• Testing the model for denoising

cd ./denoise/
python training_TF.py

• Modify training_TF.py istrain=False modeltypelst = ['FBuncertainty_f_twostage'] Replacing "testGTpath" with the path of the grount-truth images in the testing data. Replacing "modelpath1" with the path of the model trained in the first stage. Replacing "modelpath" with the path of the model trained in the second stage.

All training and test data involved in the experiments are publicly available datasets and can be downloaded from https://publications.mpi-cbg.de/publications-sites/7207/

The pretrained models can be downloaded from https://pan.baidu.com/s/1IqedS62A02THpRdvgyJEMg?pwd=rvcf

You can use our online demo in feedbackfmir.fdudml.cn

To use our demo in your local environment, you need：

• python 3.8
• gradio == 3.40.1

Load exampledata(optional) and model weight in ./exampledata and ./experiment respectively. You need to modify the corresponding code in app.py and model/fmirmodel.py to ensure that the program runs normally.

Finally, you can use the gradio demo by running：

python app.py