Space Selfies Level Up-AI Learns from the Best in the Universe

Link

Abstract and 1 Introduction

2 ways

2.1. General -Idea and 2.2. Encoder-Decoder architecture

2.3. Transformers for image restoration

2.4. Details on implementation

3 data and 3.1. Hst dataset

3.2. Galsim dataset

3.3. Jwst dataset

4 test results in JWST and 4.1. PSNR and SSIM

4.2. Visual inspection

4.3. Restoring morphological parameters

4.4. Restoring photometric parameters

5 Applications to real HST images and 5.1. Restoring single-epoch images and comparing multi-epoch images

5.2. Restoring multi-epoch HST photos and comparing multi-epoch images jwst

6 limits

6.1. The deterioration in the quality of restoration due to high noise levels

6.2. Point Source recovery test

6.3. Artifacts Due to the Pixel Relationship

7 Conclusions and Recognition

Appendix: A. Image Restoration Test with blank images only in noise

References

3. data

The purpose of the current study is to develop a transformer model that improves the quality of HST images in JWST. Thus, our training data set consists of pairs of JWST images with quality ground truth (GT) and their low -quality versions (LQ). We use a transfer learning strategy through the first pre-training our model with GT-LQ pairs of simplified expanse images and activates the pre-exemplary model with GT-LQ pairs of realistic expanse images. Therefore, we prepare two sets of training datasets.

The GT image of pre-training dataset is formed along with Galsim (Rowe et al. 2015) while its LQ counterpart is obtained only by slowing down the GT image resolution with HST point spread function (PSF) and increasing Gaussian noise. Galsim input parameters should not be unreasonable but should be carefully selected to represent the structural features of real galaxies. This is achieved by harvesting the true images of space from the HST archive and measuring their property.

The GT image of the Finetuning Dataset is obtained by collecting Galaxy images from the JWST Archive. Following the same procedure in the generation of pre-training dataset, we created their LQ counterparts. Below we described in detail the HST, Galsim, and JWST Datasets.

3.1. Hst dataset

We collect multi-band galaxy images from the following HST/ACS deep fields: The Hubble Ultra Deep Field 12 (HUDF12), HUDF-Parallel 2 (HUDFP2)[2] .[3] (Illlingworth et al. 2016; Whitaker et al. 2019) fields. F606W, F775W, and F814W filter are used from HUDF12. For HUDFP2, only F814W was selected. We used F606W and F814W from Goods-North & South.

The sexractor (Bertin & Arnouts 1996) was used to detect resources by finding five connected pixels 1.5 times above the noise level of RMS. We deleted the image of the AW × W Pixel postage-stamp centered on each resource. Image size w is set to 2[4a+5]where a sextractor's semi-major axis output and the operation [x] represents the greatest integer of no more than x. We discarded the source if

1. Its image size is greater than 128 picsels,

2. Either semi-minor axis or half light radius is smaller than the fwhm of HST/ACS PSF,

3. Its S/N value is less than 20,

4. The maximum amount of its neighboring resources is greater than the 0.8 times worth of its peak, or

5. The gradient within the middle 10 × 10 picses is sharp than the HST PSF.

The first and second criteria removes excessively large or small resources. The third condition removes extremely fuzzy resources. The fourth condition does not include cases where similar bright galaxies are present inside the seal seal image. The final condition is required because the HST automatic cosmic ray detection/removal of the pipeline often fails near the bright peaks of the stars or galaxies. In addition to these standards, we clearly identified and removed the stars using their unique locus in the size-magnitude diagram. We get a total of 66,092 seal-stamp images. Figure 3 shows the distribution of HST postage-stamp images before W = 128 cut.

As mentioned above, our pre-training dataset is created by galsim using the properties derived from our HST dataset. We choose to identify HST's galaxies through their best fit elliptical sersic parameters and flux. Sersic's circular profile is defined as

Where N represents the Sersic index, the R is the radius, the R50 is the half -light radius, and I (R50) are the surface light on the R50. The elliptical sersic profile is implemented by replacing R with the following:

where A and B indicate semi-major and semi-minutes axes, respectively.

Due to the rapid change in I (R) close to the middle (R = 0), the review of I (R) only in the discrete positions (i.e., in the middle of each pixel) leads to a significant difference between the rendered and true images of the galaxy. To reduce this, we first oversampled I (R) with ten times higher resolution in the central region (R <5 pixel) and then applied 10 × 10 nearest neighbor-averaging to provide the galaxy. The image of this sersic galaxy is combined with the HST PSF and then fits into the HST Galaxy image. Fluxes are measured by calling backgroundsubtracted pixel values ​​from HST Galaxy images.