CCNN-CGH

Real-time CGH using CCNN (zero-padding version)

Real-time end-to-end CGH network with average PSNR more than 30dB in DIV2K valitaion dataset. Compared with HoloNet and holo-encoder, we achieve the fasted speed and the best quality using compact CCNN. Moreover, CCNN-CGH is a 4K capable network and mini CCNN-CGH is the first 4K real-time network！The following tests run using RTX 3080.

1920 performance

4k performance

paper: https://doi.org/10.1109/tvcg.2023.3239670 (If it's useful, consider cite our paper!)

0. Contents

• 1, Python code and pretrained models (30 loops in DIV2K training dataset) for different networks

  xxxtrain.py will train corresponding network.

  xxxload.py will load trained model and test it on a single picture, return average generation time, PSNR, SSIM, simulated reconstruction.

  xxxpsnrssim.py will test the trained model using 100 samples of DIV2K validation dataset, return average PSNR, SSIM.

  In reorganized codes, some issues are fixed. You can change the ASM version

• 2,Captured results

  including captured videos, images and corresponding CGHs. If you have the same devices as ours, you can use the CGHs to reproduce our experiments.

1. Set up conda environment

conda create -n ccnncgh python=3.9
conda activate ccnncgh
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
pip install opencv-python
pip install tqdm
pip install scipy
pip install scikit-image

The environment has been tested in Windows 10 and Ubuntu 20.04 in 2022.7. We use python 3.9, Pytorch 1.10.

2. Run CCNN-CGH

For better compatibility with complex values, we use complexPyTorch from https://github.com/wavefrontshaping/complexPyTorch. Their new version should also work.

Change the file path in the code.

Run corresponding python files. CCNN uses 4init, mini CCNN uses 2init.

3.Comparision with other networks

• Comparision with HoloNet

  Before running HoloNet, make sure your GPU has more than 10GB memory.

  HoloNet and some codes are from https://github.com/computational-imaging/neural-holography

  The U-Net used in HoloNet is from https://github.com/vsitzmann/pytorch_prototyping

• Comparision with Holo-encoder

  Holo-encoder is from:https://github.com/THUHoloLab/Holo-encoder

  For pytorch version, we use https://github.com/flyingwolfz/holoencoder-python-version

• Comparision with tensor holography

  We run tensor holography in Ubuntu 20.04 using their code and pretrained model from https://github.com/liangs111/tensor_holography

  Matlab code is used for simulation. See ASM: https://github.com/flyingwolfz/angular-spectrum-method