Transformer-Based Restoration: Quantitative Gains and Boundaries in Space Data

cover
7 May 2025

Authors:

(1) Hyosun park, Department of Astronomy, Yonsei University, Seoul, Republic of Korea;

(2) Yongsik Jo, Artificial Intelligence Graduate School, UNIST, Ulsan, Republic of Korea;

(3) Seokun Kang, Artificial Intelligence Graduate School, UNIST, Ulsan, Republic of Korea;

(4) Taehwan Kim, Artificial Intelligence Graduate School, UNIST, Ulsan, Republic of Korea;

(5) M. James Jee, Department of Astronomy, Yonsei University, Seoul, Republic of Korea and Department of Physics and Astronomy, University of California, Davis, CA, USA.

Abstract and 1 Introduction

2 Method

2.1. Overview and 2.2. Encoder-Decoder Architecture

2.3. Transformers for Image Restoration

2.4. Implementation Details

3 Data and 3.1. HST Dataset

3.2. GalSim Dataset

3.3. JWST Dataset

4 JWST Test Dataset Results and 4.1. PSNR and SSIM

4.2. Visual Inspection

4.3. Restoration of Morphological Parameters

4.4. Restoration of Photometric Parameters

5 Application to real HST Images and 5.1. Restoration of Single-epoch Images and Comparison with Multi-epoch Images

5.2. Restoration of Multi-epoch HST Images and Comparison with Multi-epoch JWST Images

6 Limitations

6.1. Degradation in Restoration Quality Due to High Noise Level

6.2. Point Source Recovery Test

6.3. Artifacts Due to Pixel Correlation

7 Conclusions and Acknowledgements

Appendix: A. Image restoration test with Blank Noise-Only Images

References

7. CONCLUSIONS

We have showcased astronomical image restoration from HST quality to JWST quality using the efficient Transformer model via transfer learning. The pretraining dataset was created by rendering GT galaxy images based on analytic profiles and generating corresponding LQ versions by reducing their resolution and introducing noise. The finetuning dataset was produced by sampling GT galaxy images from deep JWST images, which were then degraded to the LQ images in a similar fashion.

With the test dataset, we find that the restored images show significantly enhanced correlations with the GT images than their original LQ versions, reducing the scatters of isophotal photometry, Sersic index, and halflight radius by factors of 4.4, 3.6, and 4.7, respectively, with Pearson correlation coefficients approaching unity. We also visually confirm that the restored images are superior in terms of resolution and noise level. When we applied our model to real low-exposure HST images, the restored images also show significantly improved correlations with their multi-exposure versions, although the absence of their real GT images limits our interpretations.

We discuss a few limitations of our model. First, the performance degrades in high noise regimes, where the background rms approaches ∼10% of the object peak values. Second, highly correlated noise can be misinterpreted as astronomical features, leading to the manifestation of low-surface brightness features. Third, the restoration of point sources is less than optimal.

Although it is possible to further improve the model with larger training datasets and enhanced training strategies, we anticipate that our current Transformer-based deep learning model will prove useful for a number of scientific applications, including precision photometry, morphological analysis, and shear calibration.

ACKNOWLEDGEMENTS

This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub). M.J.J. acknowledges support for the current research from the National Research Foundation (NRF) of Korea under the programs 2022R1A2C1003130 and RS-2023-00219959. This work is based [in part] on observations made with the NASA/ESA/CSA James Webb Space Telescope. The data were obtained from the Mikulski Archive for Space Telescopes at the Space Telescope Science Institute, which is operated by the

Figure 11. Performance degradation cases. As the noise level increases, the restoration depends on fewer and fewer pixels. Thus, more frequent degradation cases occur with higher noise levels. Shown here are examples of loss of spiral arms (first row), incorrect ellipticity (second and third row), and failed de-blending of multiple peaks (fourth row).

Figure 12. Point source restoration test. We created 1000 LQ-GT pairs of point source images and compared the mean RS image with the mean GT image (see text for details). The profile in the RS image is slightly more extended than that in the GT image.

Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-03127 for JWST.

Facilities: JWST (NIRCam), HST (ACS)

Software: numpy (Harris et al. 2020), scipy (Virtanen et al. 2020), matplotlib (Hunter 2007), astropy (Astropy Collaboration et al. 2013, 2018), photutils (Bradley et al. 2023), SExtractor (Bertin & Arnouts 1996), GalSim (Rowe et al. 2015)

Figure 13. Application to HST galaxies sampled from multi-epoch drizzled images. Since the drizzling algorithm induces substantial noise correlation, the input data have different noise properties from our training datasets. The RS images show that the correlated noise creates some low-surface brightness artifacts in the outskirts.

APPENDIX

A. IMAGE RESTORATION TEST WITH BLANK NOISE-ONLY IMAGES

One of the key requirements of our deep-learning-based restoration model is that the model should not generate any false object images by overinterpreting the noise when the image contains no real astronomical source. To test this, we created blank noise-only images by varying the random seed and the noise level. We used 10 different random seeds, and for each random seed, we generate 1000 images, where the mean and standard deviation of the noise were set to

Figure 14. Noise-only image restoration test. (a) Examples of blank image restoration. The rms noise increases from top to bottom. No pseudo-object is produced while the noise is significantly reduced. (b) Test with different random seeds. Each data point represents statistics from 1,000 images. The mean and standard deviation of the RS images are consistent across different random seeds. (c) Sames as (b) but for different noise levels. Regardless of the input (LQ) image’s noise level, the output (RS) image’s background mean and noise level stay consistently low.

mimic those of a randomly selected galaxy image from the JWST train dataset. The RS images created from these blank images were carefully inspected. No image was found to contain pseudo-sources. Figure 14 presents the test results.

REFERENCES

Akhaury, U., Starck, J.-L., Jablonka, P., Courbin, F., & Michalewicz, K. 2022, Frontiers in Astronomy and Space Sciences, 9, 357, doi: 10.3389/fspas.2022.1001043

Alsaiari, A., Rustagi, R., Alhakamy, A., Thomas, M. M., & Forbes, A. 2019, in 2019 IEEE 2nd International Conference on Information and Computer Technologies (ICICT), 126–132, doi: 10.1109/INFOCT.2019.8710893

Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33, doi: 10.1051/0004-6361/201322068

Astropy Collaboration, Price-Whelan, A. M., Sip˝ocz, B. M., et al. 2018, AJ, 156, 123, doi: 10.3847/1538-3881/aabc4f

Bertin, E., & Arnouts, S. 1996, A&AS, 117, 393, doi: 10.1051/aas:1996164

Bradley, L., Sip˝ocz, B., Robitaille, T., et al. 2023, astropy/photutils: 1.10.0, 1.10.0, Zenodo, doi: 10.5281/zenodo.1035865

D´ıaz Baso, C. J., de la Cruz Rodr´ıguez, J., & Danilovic, S. 2019, A&A, 629, A99, doi: 10.1051/0004-6361/201936069

Elhakiem, A. A., Elsaid Ghoniemy, T., & Salama, G. I. 2021, in 2021 Tenth International Conference on Intelligent Computing and Information Systems (ICICIS), 51–56, doi: 10.1109/ICICIS52592.2021.9694140

Ellis, R. S., McLure, R. J., Dunlop, J. S., et al. 2013, ApJL, 763, L7, doi: 10.1088/2041-8205/763/1/L7

Goodfellow, I., Pouget-Abadie, J., Mirza, M., et al. 2014, Advances in neural information processing systems, 27

Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357, doi: 10.1038/s41586-020-2649-2

Hunter, J. D. 2007, Computing in Science and Engineering, 9, 90, doi: 10.1109/MCSE.2007.55

Illingworth, G., Magee, D., Bouwens, R., et al. 2016, arXiv e-prints, arXiv:1606.00841, doi: 10.48550/arXiv.1606.00841

Jones, R., & Wykes, C. 1989, Holographic and Speckle Interferometry (Cambridge University Press)

Kalele, G. 2023, in 2023 1st International Conference on Innovations in High Speed Communication and Signal Processing (IHCSP), 212–215, doi: 10.1109/IHCSP56702.2023.10127178

Koekemoer, A. M., Ellis, R. S., McLure, R. J., et al. 2013, ApJS, 209, 3, doi: 10.1088/0067-0049/209/1/3

Krizhevsky, A., Sutskever, I., & Hinton, G. E. 2012, Advances in neural information processing systems, 25

Lanusse, F., Mandelbaum, R., Ravanbakhsh, S., et al. 2021, MNRAS, 504, 5543, doi: 10.1093/mnras/stab1214

Ledig, C., Theis, L., Huszar, F., et al. 2016, arXiv e-prints, arXiv:1609.04802, doi: 10.48550/arXiv.1609.04802

Liu, M.-Y., Huang, X., Yu, J., Wang, T.-C., & Mallya, A. 2021, Proceedings of the IEEE, PP, 1, doi: 10.1109/JPROC.2021.3049196

Loshchilov, I., & Hutter, F. 2018, Fixing Weight Decay Regularization in Adam. https://openreview.net/forum?id=rk6qdGgCZ

Lucy, L. B. 1974, AJ, 79, 745, doi: 10.1086/111605

Rajeev, R., Samath, J., & Karthikeyan, N. 2019, Journal of Medical Systems, 43, doi: 10.1007/s10916-019-1371-9

Richardson, W. H. 1972, Journal of the Optical Society of America (1917-1983), 62, 55

Ronneberger, O., Fischer, P., & Brox, T. 2015, in Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, Springer, 234–241

Rowe, B., Jarvis, M., Mandelbaum, R., et al. 2015, Astronomy and Computing, 10, 121, doi: 10.1016/j.ascom.2015.02.002

Schawinski, K., Zhang, C., Zhang, H., Fowler, L., & Santhanam, G. K. 2017, MNRAS, 467, L110, doi: 10.1093/mnrasl/slx008

Shepp, L. A., & Vardi, Y. 1982, IEEE Transactions on Medical Imaging, 1, 113, doi: 10.1109/TMI.1982.4307558

Simkin, S. M. 1974, A&A, 31, 129

Sureau, F., Lechat, A., & Starck, J. L. 2020, A&A, 641, A67, doi: 10.1051/0004-6361/201937039

Sweere, S. F., Valtchanov, I., Lieu, M., et al. 2022, MNRAS, 517, 4054, doi: 10.1093/mnras/stac2437

Tikhonov, A. N., & Goncharsky, A. V. 1987, Ill-posed problems in the natural sciences (Mir Publishers)

Tran, L., Nguyen, S., & Arai, M. 2021, GAN-Based Noise Model for Denoising Real Images (Springer, Cham), 560–572, doi: 10.1007/978-3-030-69538-5 34

Tripathi, S., Lipton, Z., & Nguyen, T. 2018, arXiv preprint arXiv:1803.04477

Vaswani, A., Shazeer, N., Parmar, N., et al. 2017, in Advances in neural information processing systems, 5998–6008

Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020, Nature Methods, 17, 261, doi: 10.1038/s41592-019-0686-2

Wakker, B. P., & Schwarz, U. J. 1988, A&A, 200, 312

Wang, Z., Wang, L., Duan, S., & Li, Y. 2020, Journal of Physics: Conference Series, 1550, 032127, doi: 10.1088/1742-6596/1550/3/032127

Whitaker, K. E., Ashas, M., Illingworth, G., et al. 2019, ApJS, 244, 16, doi: 10.3847/1538-4365/ab3853

Williams, R. J., & Zipser, D. 1989, Neural computation, 1, 270

Yan, L., Fang, H., & Zhong, S. 2012, Optics Letters, 37, 2778, doi: 10.1364/OL.37.002778

Zamir, S. W., Arora, A., Khan, S., et al. 2022, in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 5728–5739

Zhang, K., Zuo, W., Chen, Y., Meng, D., & Zhang, L. 2017, IEEE Transactions on Image Processing, 26, 3142, doi: 10.1109/TIP.2017.2662206

Zhang, Q., Xiao, J., Tian, C., Lin, J., & Zhang, S. 2022, CAAI Transactions on Intelligence Technology, 8, n/a, doi: 10.1049/cit2.12110

This paper is available on arxiv under CC BY 4.0 Deed license.