This repository contains the codes for the neural reconstruction of solar images using the multiobject multiframe blind deconvolution framework. There are several codes that implement different strategies.

This code implements the MFBD method using algorithm unrolling. It uses gradient descent to optimize the MFBD loss function with respect to the wavefront coefficients. The current estimation of the coefficients and the computed gradient at each step are used by a neural network to provide an updated estimate of the gradient. This process is unrolled K times and the neural networks are trained sequentially for memory efficiency.

This is the same approach followed by Asensio Ramos & Olspert (2021) but with a much deeper encoder and changing the recurrent neural network with a convolutional neural network acting in time.

This code implements the classic MFBD algorithm, using the wavefronts as input and using PyTorch for computing the derivatives. It can be used as a testbed for implementing new ideas and also as a comparison for the results of the previous codes.

The aim of this code is to train an autoencoder to reconstruct the point spread function and use it in the previous codes, instead of using wavefronts. The idea is based on the fact that the relation between wavefronts and point spread functions is non unique in a point symmetric pupil. For instance, a change of sign of all even modes in the Zernike basis produces exactly the same PSF. This can be confusing for the neural network. If we train an autoencoder to compress PSFs into a latent space, this ambiguity is removed. The inverse process carried out by the previous methods will then be carried out in the latent space. [WIP]

If you want to train the models yourself, you can find the training data here. It is a tarfile containing a file in zarr format for the Swedish Solar Telescope, containing observations for 3934 A observed with CHROMIS, as well as 8542 A and 6173 A, both observed with CRISP.

https://cloud.iac.es/index.php/s/xiYP7j4rxRk4TkH

You need to install the following dependencies:

numpy 
opencv 
ffmpeg 
matplotlib 
tqdm 
scipy 
astropy 
sunpy 
pillow 
numba 
zarr 
patchify 
configobj 
millify 
skimage 
h5py 
telegram 
nvidia-ml-py3
patchify
pytorch

They can be installd in a new environment using

conda create --name mfbd python=3.9
conda activate mfbd
conda install -c conda-forge numpy opencv ffmpeg matplotlib tqdm scipy astropy sunpy pillow numba zarr patchify configobj millify skimage h5py telegram pytorch
conda install -c fastai nvidia-ml-py3
pip install millify patchify

pytorch can be installed following the instructions in https://pytorch.org, but they probably reduce to:

conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia

During training, you can have notifications showing how the process is goind via Telegram. For this, look for your Telegram token, create a new chat. This is done by opening a new conversation with @botfather and typing /newbot. You will get a new token that you need to add to your environment variables as shown below. Every chat has a chat ID, and the code needs this chat ID to send Telegram messages using Python. It can be found by sending a message to the bot and then running:

import requests
TOKEN = "YOUR TELEGRAM BOT TOKEN"
url = f"https://api.telegram.org/bot{TOKEN}/getUpdates"
print(requests.get(url).json())

One of the outputs of the JSON file will be the chat ID. Add that to the environment variables.

export TELEGRAM_TOKEN='your_token'
export TELEGRAM_CHATID='your_chat_id'