Skip to content

aasensio/learned_mfbd

Repository files navigation

Real-time multiframe blind deconvolution of solar images (submitted to Astronomy & Astrophysics)

Introduction

The quality of images of the Sun obtained from the ground are severely limited by the perturbing effect of the turbulent Earth's atmosphere. The post-facto correction of the images to compensate for the presence of the atmosphere require the combination of high-order adaptive optics techniques, fast measurements to freeze the turbulent atmosphere and very time consuming blind deconvolution algorithms. Under mild seeing conditions, blind deconvolution algorithms can produce images of astonishing quality. They can be very competitive with those obtained from space, with the huge advantage of the flexibility of the instrumentation thanks to the direct access to the telescope. In this contribution we leverage deep learning techniques to significantly accelerate the blind deconvolution process and produce corrected images at a peak rate of ~100 images per second. We present two different architectures that produce excellent image corrections with noise suppression while maintaining the photometric properties of the images. As a consequence, polarimetric signals can be obtained with standard polarimetric modulation without any significant artifact.
With the expected improvements in computer hardware and algorithms, we anticipate that on-site real-time correction of solar images will be possible in the near future.

Dependencies

This code depends on PyTorch >=0.4.0 but can be easily adapted to prior versions with not much effort.

Usage

We provide a simple example of a monochromatic H-alpha burst observation and two scripts to use the two neural networks trained for this work. We note that observations need to be multiplied by 0.001 before entering the networks and multiplied by 1000 after the exit. The input should be given in counts as output by the data reduction pipeline of CRISP@SST.

python test_recurrent.py
python test_encdec.py

Training

We do not provide the training scripts because the data used for training is too heavy. They should be pretty straighforward to build following the standard PyTorch training approach.

About

Learned MFBD

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages