This dataset contains raw and processed astronomical CCD observations from the MuSCAT instruments, organized for deep-sky image denoising and related low-level image restoration research.
The repository structure follows the organization used in the Denoising-Deep-Sky project, where data are grouped by instrument, band, and observation session.

The data are intended primarily for:

astronomical image denoising
raw-to-clean image restoration
CCD noise modeling
calibration-aware image processing
benchmarking denoising methods on real telescope data

Dataset Structure

The dataset is organized by instrument, spectral band, and observation session.

A typical directory layout is:

dataset_root/
├── muscat3_g/
│   ├── <observation_id_1>/
│   │   ├── calib/
│   │   │   ├── BIAS_*.fits.fz
│   │   │   ├── BPM_*.fits.fz
│   │   │   ├── DARK_*.fits.fz
│   │   │   └── SKYFLAT_*.fits.fz
│   │   ├── data/
│   │   │   ├── <obs_id>_calib.fits.fz
│   │   │   ├── <obs_id>_mask.fits.fz
│   │   │   ├── <obs_id>_mean.fits.fz
│   │   │   ├── <obs_id>_os.fits.fz
│   │   │   └── <obs_id>_raw.fits.fz
│   │   ├── e91/
│   │   ├── calib.json
│   │   └── info.json
│   └── <observation_id_2>/
│       └── ...
└── muscat4_i/
    └── ...

Data Contents

Each observation directory may include the following components.

Calibration files (`calib/`)

BIAS_*.fits.fz
Master bias frame. Represents the electronic offset estimate.
BPM_*.fits.fz (Empty) Bad pixel map.
DARK_*.fits.fz
Master dark-current template.
SKYFLAT_*.fits.fz
Sky flat used for pixel-response and vignetting correction.

Observation data (`data/`)

<obs_id>_raw.fits.fz
Raw CCD frame.
<obs_id>_os.fits.fz
Overscan-corrected frame, typically used as the noisy input.
<obs_id>_calib.fits.fz
Calibrated frame.
<obs_id>_mean.fits.fz
High-SNR stacked reference frame built from multiple exposures.
<obs_id>_mask.fits.fz
Background/source mask.

Additional files

e91/
Additional calibrated frames used to construct stacked references.
calib.json
Calibration-related metadata.
info.json
Observation/session metadata.

Data Fields

This repository does not follow a tabular row-based structure.
Instead, each sample is stored as a set of FITS files and JSON metadata under one observation directory.

Important file-level fields include:

instrument: e.g. MuSCAT-3 or MuSCAT-4
band: e.g. g, r, i
observation_id: identifier for one observation session
raw frame: original CCD image
overscan-corrected frame: original CCD image - overscan
calibrated frame: processed image using calibration products
mean frame: stacked high-SNR reference image
mask: source/background mask (background = 1, source = 0)
calibration metadata: gain, dark, bias, flat-related information when available

Supported Tasks and Use Cases

This dataset is suitable for:

real-world astronomical image denoising
supervised image-to-image restoration
physics-based CCD noise modeling
calibration-aware low-level vision
benchmarking telescope image restoration pipelines

Possible training pairs include:

original pairs: *_os.fits.fz ----> *_calib.fits.fz
synthesized pairs: *_syn.fits.fz ----> *_mean.fits.fz

Depending on the experiment design, users may also incorporate:

*_raw.fits.fz
*_calib.fits.fz
calibration frames from calib/
masks from *_mask.fits.fz

Loading the Data

Because the dataset is distributed as FITS files instead of a standard row-based file format such as CSV or Parquet, it is generally intended to be accessed as files from the repository rather than directly through the default Hugging Face dataset viewer.

Example with Python:

from astropy.io import fits

path = "muscat3_g/<observation_id>/data/<obs_id>_raw.fits.fz"

with fits.open(path) as hdul:
    image = hdul[1].data
    header = hdul[1].header

print(image.shape)
print(header)

Citation

If you use this dataset, please cite the original project:

@misc{liu2025denoisingdeepsky,
  title={Denoising the Deep Sky: Physics-Based CCD Noise Formation for Astronomical Imaging},
  author={Shuhong Liu and Xining Ge and Ziying Gu and Quanfeng Xu and Ziteng Cui and Lin Gu and Xuangeng Chu and Jun Liu and Dong Li and Tatsuya Harada},
  year={2025},
  eprint={2508.08079},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}