Skip to main content
SearchLoginLogin or Signup

Neural Image Compression on Solar Dynamics Observatory

Presentation #125.21 in the session General Topics: Solar — Poster Session.

Published onOct 20, 2022
Neural Image Compression on Solar Dynamics Observatory

Recently end-to-end optimized artificial neural networks (ANN) have shown strong potential in the field of image compression. They are capable of outperforming conventional hand-engineered algorithms for lossy and lossless image compression. We propose a method to use ANN to reduce the amount of information needed to be stored and retrieved in Space missions studying solar dynamics. Solar Dynamics Observatory (SDO) mission gathers 1.4 Tera-bytes of data each day orbiting in space. To save and archive this huge amount of data we are proposing ad-hoc and application-specific image compression algorithms. By training individual neural networks to compress all variety of images downloaded from SDO spacecraft, we can easily propose algorithms for compressing images that are outperforming the popular image compression algorithms such as JPEG. The most important advantage of compressing images using neural networks, in comparison to conventional methods, is that we directly optimize Rate-Distortion criterion. By enforcing the target bitrate which is defined by the capacity of the channel or the available amount of storage for archiving each image, we can optimize the parameters of the artificial neural network to reach the targeted bit rate. This job is being done by optimizing the trade-off between rate and distortion, in which the distortion is measured by a distance metric between the uncompressed image and its corresponding reconstruction. To make the distribution of reconstructions and uncompressed images closer to each other, we adopted adversarial training in the framework of generetive adversarial network (GAN) as well. These neural networks could be trained on smaller crops of images to reach a targeted bit rate, and then be used to compress images at any given size to that bit rate. Even if different resolutions of the sun observations need to be accessible to the solar community for scientific purposes, we train only one neural network to compress images of different sizes at the desired quality.


No comments here