-
Notifications
You must be signed in to change notification settings - Fork 11
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Showing
2 changed files
with
122 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,35 @@ | ||
@ARTICLE{deepcr, | ||
author = {{Zhang}, K. and {Bloom}, J.~S.}, | ||
title = "{deepCR: Cosmic Ray Rejection with Deep Learning}", | ||
journal = {arXiv e-prints}, | ||
archivePrefix = "arXiv", | ||
eprint = {1907.09500}, | ||
primaryClass = "astro-ph.IM", | ||
keywords = {Astrophysics - Instrumentation and Methods for Astrophysics, Computer Science - Computer Vision and Pattern Recognition}, | ||
year = 2019, | ||
month = jul, | ||
adsurl = {https://ui.adsabs.harvard.edu/abs/2019arXiv190709500Z}, | ||
adsnote = {Provided by the SAO/NASA Astrophysics Data System} | ||
} | ||
|
||
@ARTICLE{lacosmic, | ||
author = {{van Dokkum}, P.~G.}, | ||
title = "{Cosmic-Ray Rejection by Laplacian Edge Detection}", | ||
journal = {\pasp}, | ||
eprint = {astro-ph/0108003}, | ||
keywords = {Instrumentation: Detectors, Methods: Data Analysis-techniques: image processing}, | ||
year = 2001, | ||
month = nov, | ||
volume = 113, | ||
pages = {1420-1427}, | ||
doi = {10.1086/323894}, | ||
adsurl = {https://ui.adsabs.harvard.edu/abs/2001PASP..113.1420V}, | ||
adsnote = {Provided by the SAO/NASA Astrophysics Data System} | ||
} | ||
|
||
@inproceedings{pytorch, | ||
title = {Automatic differentiation in PyTorch}, | ||
author = {Paszke, Adam and Gross, Sam and Chintala, Soumith and Chanan, Gregory and Yang, Edward and DeVito, Zachary and Lin, Zeming and Desmaison, Alban and Antiga, Luca and Lerer, Adam}, | ||
booktitle = {NIPS-Workshop}, | ||
year = {2017} | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,87 @@ | ||
--- | ||
title: 'deepCR: Cosmic Rejection with Deep Learning' | ||
tags: | ||
- Python | ||
- Pytorch | ||
- astronomy | ||
- image processing | ||
- cosmic ray | ||
- deep learning | ||
|
||
authors: | ||
- name: Keming Zhang | ||
orcid: 0000-0002-9870-5695 | ||
affiliation: 1 # (Multiple affiliations must be quoted) | ||
- name: Joshua S. Bloom | ||
orcid: 0000-0002-7777-216X | ||
affiliation: "1, 2" | ||
affiliations: | ||
- name: Department of Astronomy, University of California, Berkeley | ||
index: 1 | ||
- name: Lawrence Berkeley National Laboratory | ||
index: 2 | ||
date: 5 August 2019 | ||
bibliography: paper.bib | ||
--- | ||
|
||
# Summary | ||
|
||
Astronomical imaging and spectroscopy data are frequently corrupted by | ||
"cosmic rays" (CR) which are high energy charged particles that are instrumental, | ||
terrestrial, or cosmic in origin. When such particles pass through solid state | ||
detectors, such as charged coupled devices (CCDs), they create excess | ||
flux in the pixels hit which lead to artifacts in images. These | ||
artifacts must be identified and either masked or replaced, before | ||
further scientific analysis could be done on the image data. It is straightforward | ||
to identify these artifacts when multiple exposures of the same field are | ||
taken. In such cases, a median image could be calculated from aligned single | ||
exposures, effectively creating a CR-free image. Each one of the exposures | ||
is then compared with the median image to identify the cosmic rays. However, | ||
when CCD read-out times are non-negligible, or when sources of | ||
interest are transient or variable, cosmic ray rejection with multiple | ||
exposures can be sub-optimal or infeasible. These cases would require specialized | ||
algorithms to detect cosmic rays in single images. | ||
|
||
``deepCR`` is a Python package for single frame cosmic ray rejection which is | ||
based on deep learning and written with the Pytorch framework [@pytorch]. | ||
Since ``deepCR`` is based on deep learning, different models trained on | ||
data taken with different instrument configurations are required, when applied to different | ||
data. The current version of ``deepCR`` is prepackaged with model for Hubble | ||
Space Telescope ACS/WFC imaging data, and we expect models available | ||
to grow with contribution from the community. We plan to host a "model zoo" | ||
which enables ``deepCR`` to work across different instrument configurations. | ||
|
||
The API of ``deepCR`` includes functionality for both applying models and | ||
training models. To apply an available model, ``deepCR`` takes in an input image | ||
and produces a cosmic ray mask and an "inpainted" image, with | ||
the artifact pixels replaced with ``deepCR`` predictions. To train a new model, | ||
users would feed in custom dataset to the training API, which is automated. | ||
``deepCR`` works with both CPU, which is well-threaded at application time, and GPU. | ||
On GPU, training a new model takes as short as 20 minutes, | ||
while applying ``deepCR`` on a 10 Mpix image requires less than 0.2 second, | ||
orders of magnitude faster than current state of the art ``LACosmic`` [@lacosmic]. | ||
![Example of cosmic ray contaminated image cutouts (first row), deepCR | ||
cosmic ray mask predictions (middle row), and original image with artifact | ||
pixels replaced with deepCR predictions (last row).](imgs/postage-sm.png) | ||
|
||
|
||
In the paper accompanying ``deepCR`` [@deepcr], the authors showed that | ||
on Hubble Space Telescope (HST) ACS/WFC data, | ||
``deepCR`` is more robost, and at least as fast as the current | ||
state-of-the-art single frame cosmic ray rejection package, ``LACosmic``. The API | ||
of ``deepCR`` serve as a drop in replacement for ``LACosmic``, | ||
so that users may experiment with different packages easily. At | ||
reasonable false detection rates, ``deepCR`` achieved near perfect | ||
cosmic ray detection in extragalactic and globular cluster fields, and above | ||
90% in more difficult dense stellar fields in nearby resolved galaxies. | ||
Since HST imaging is among the hardest cosmic ray rejection to be | ||
solved, ``deepCR`` would work well across many different instrument set-ups, | ||
including ground based imaging and spectroscopy. The combination of | ||
speed and accuracy of ``deepCR`` allows astronomers to potentially save | ||
large amounts of precious observational and computational resources. | ||
|
||
# Acknowledgements | ||
|
||
This work was supported by a Gordon and Betty Moore Foundation Data-Driven Discovery grant. | ||
|
||
# References |