Even Amnesty International needs glasses: when space images become very mysterious to fix them
Authors:
(1) Hyosun Park, Department of Astronomy, Yonsei University, Seoul, Republic of Korea;
(2) Yongsik Jo, Graduate Studies School of AI, ULSAN, Korea Republic;
(3) Seokun Kang, College of Graduate Studies for Artificial Intelligence, UNIST, Ulsan, Republic of Korea;
(4) Tahwan Kim, College of Graduate Studies, Artificial Intelligence, UNIST, Ulsan, Republic of Korea;
(5) M. James G, Astronomy Department, University of Uniony, Seoul, Republic of Korea and the Department of Physics and Astronomy, University of California, Davis, California, USA.
Links table
Abstract and 1 introduction
2 method
2.1. Overview and 2.2. Architecture coding encryption
2.3. Transformers to recover pictures
2.4. Implementation details
3 data and 3.1. HST data collection
3.2. GALSIM Data collection
3.3. JWST Data collection
4 JWST and 4.1 testing collection results. PSNR and SSIM
4.2. Optical inspection
4.3. Restore morphological parameters
4.4. Restore optical parameters
5 app on real HST and 5.1 images. Restoring single images and comparison with multi -episode images
5.2. Restoring multi -EPOCH and comparison with multi -jwst images
6 restrictions
6.1. Downtown in the quality of restoration due to the high level of noise
6.2. Point source recovery test
6.3. The artifacts due to the bonding of the pixels
7 conclusions and estimates
Approach: A. Images recovery test with empty noise images
Reference
6. Restrictions
6.1. Downtown in the quality of restoration due to the high level of noise
Although our transformer’s deep -based learning model provides a recent performance in both precision and reduction in moderate noise levels, the appropriate restoration is inevitable, becomes impossible when the noise level exceeds a specific threshold.
Figure 11 shows four such examples. The upper row explains the condition, as the distinctive spiral arm is not restored in the image of GT. When checking the corresponding LQ image, we believe that the loss of information resulting from noise is very important that it cannot be hurt to the presence of a spiral arm. The second row shows a case, as the angle of the GT image position is not properly restored. Again, we doubt that the LQ image is very noisy to enable the appropriate reasoning for the location of the GT image. The third row is an example, as the elliptical for the RS image is much retrieving from the GT image. In the last row, we clarify a case, as the deep learning model failed to unveil the distinct atoms of the GT image.
Since the exact definition of the failure of the restoration is a personal matter, it is difficult to quote a precise threshold. However, from visual inspections, we suggest that the frequency of the failure increases significantly when the RMS value (after the LE-MAX normalization) is larger than ∼0.1.
6.2. Point source recovery test
In principle, the perfect Deconvolution algorithm should be regained by the point of point to the Delta function, which is smaller without boundaries of the pixels. In the traditional installation in the field of Fourieh, it is a difficult task because the process is numbers unstable. The resulting images often show many resonance effects around a bright central peak.
Since we excluded the stars from the training data set, our deep learning model has not explicitly learned to remove the source of the point source. Thus, it is interesting to study the quality of the deep learning model, which has been trained in galaxy images only, and the restoration of points. We take the point of the point of the point of point as follows. First, we created 1000 GWST image. Since we have not removed JWST PSF from the training data set, Star GT photos should not resemble the functions of the Delta but JWST PSF. We performed this by connecting one pixel with Gausi that corresponds to the size of Kerneel with JWST PSF. Note that we are randomly randomly with the stars inside the central 24 x 24 pixels of 64 x 64 photos. After that, we created their LQ versions by further selecting GT images with HST PSF and adding noise. Finally, these LQ images are restored by our deep learning form. To investigate the systematic effect, we draw pictures of 1000 GT and RS separately after aligning its centers.
Figure 12 shows that the difference is accurate when we compare the chimneys (left) and RS visually. However, the remaining image (right) shows that PSF is systematically larger RS. Thus, we conclude that our deep learning model, which was only trained with Galaxy images, performs less than perfect for points.
6.3. The artifacts due to the bonding of the pixels
In our generation of LQ pictures, we assume that noise is Russian. However, in real astronomical images, especially when we create deep images by accumulating many hanging exposters, there are great connections to the interpixel noise. We find that these noise links between the pixels create an uncharacty artifact.
Figure 13 displays some examples of these artifacts. LQ samples are taken here from multiple terrible images. The appearance of attached noise
Even from visual inspection. RS images show that associated noise creates some low -surface surface artifacts on the outskirts of the galaxy, which are absent in JWST photos. Treating this problem can include strategies such as the use of a different nucleus to prepare for images or benefit from the most advanced deep learning algorithms. Explore these solutions will be a major axis for our future work.
