Skip to content

Latest commit

 

History

History
112 lines (99 loc) · 5.22 KB

paper.md

File metadata and controls

112 lines (99 loc) · 5.22 KB
Raw
title tags authors affiliations date bibliography aas-doi aas-journal
deepCR: Cosmic Rejection with Deep Learning
Python
Pytorch
astronomy
image processing
cosmic ray
deep learning
name orcid affiliation
Keming Zhang
0000-0002-9870-5695
1
name orcid affiliation
Joshua S. Bloom
0000-0002-7777-216X
1, 2
name index
Department of Astronomy, University of California, Berkeley
1
name index
Lawrence Berkeley National Laboratory
2
15 August 2019
paper.bib
10.3847/1538-4357/ab3fa6
Astrophysical Journal

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.

Neural network architecture of deepCR. Feature maps are represented by gray boxes while the number of channels and example feature map dimensions are indicated on the top of and to the left of each feature map, respectively. Different computational operations are marked in the legend to the lower left. Unfilled boxes to the right of blue arrows represent feature maps directly copied from the left, which are to be concatenated with the adjacent feature map. To apply the inpainting model, the predicted mask (dotted box at left) is concatenated with the original image as the input.

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].

Examples of cosmic ray contaminated image cutouts (first row), deepCR cosmic ray mask predictions (middle row), and images with artifact pixels replaced with deepCR predictions (last row).

In the paper accompanying deepCR [@deepcr], the authors showed that on Hubble Space Telescope (HST) ACS/WFC data, deepCR is more robust, 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, and has made use of the following software:

astropy [@astropy]; astrodrizzle [@astrodrizzle]; numpy [@vdw11]; scipy [@scipy]; matplotlib [@matplotlib]; astroscrappy [@astroscrappy]; pytorch [@pytorch]; Jupyter [@jupyter]; Scikit-image [@scikit-image]

References