The coronagraph data calibration routine in SolarSoft employs a fuzzy algorithm to generate a level 1 FITS file, addressing missing data holes. This project endeavors to introduce a novel deep learning methodology, leveraging Partial Convolutions, to effectively address typical manifestations of data loss in coronagraphic imagery. By harnessing the capabilities of Partial Convolutions, this approach aims to enhance the robustness and efficiency of data restoration processes.\
To adapt the backbone's architecture for coronagraph inpainting, understanding the nature of information loss is crucial. We deal with squared missing holes of size (32,32), which can compound into larger ones. We aim to enhance partial convolution efforts by comprehensively understanding the constraints of our problem.
-
Multi-layered UNet like Partial convolution Network: Recursive architecture designed to strike progressive locations of the image through every layer.
-
Multi-layered Fourier Space Based Variational Autoencoder: Generative model whose decoder is designed to continuesly update the mask towards the center of each loss data chunk. For faster inference and resource management, Fourier space convolutions are used as decoder and skip connections to the decoder are made through fourier space convolutions.
-
Multi-layered Fourier Space Based Partial Convolution Network: Designed a Partial Convolution UNet like network that analyses both linear spaces to extract features.
1 and 3 are trained in a Teacher Forcing fashion, as the default training method for recursive architectures. Also, they have Cross Multihead Attention in the channel dimention between:
There is a lot of relations between the Fourier Space and Convolutions:
Given that
These models will be implemented in the Python library CorKit.
pip install corkit
as part of the level_1
routines.
# Htuning
python3 -m module.<model_type>.htuning >> module/<model_type>/htuning.txt
# Training
python3 -m training fit -c module/<model_type>/config.yaml >> module/<model_type>/training.txt
This project was inspired by the work of SolarSoft developers and the researchers behind partial convolutional layers, which form the backbone of this application.