Plate-YOLO: An End-to-End Object Detection Framework for Robust Astrometric Registration of Historical Plates

arXiv:2610.25677 · astro-ph.IM · Submitted 2026-10-19 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: I'm Vera.

Jocelyn: And I'm Jocelyn, here with Subrahmanyan.

Vera: We're looking at the paper: "Plate-YOLO: An End-to-End Object Detection Framework for Robust Astrometric Registration of Historical Plates".

Jocelyn: Here's the gist of the paper. Historical astronomical plates provide a long temporal baseline for time-domain astrophysics, but degradation and structural artifacts can hinder automated astrometric calibration.

Vera: PlateYOLO is an end-to-end object detection framework that combines source detection and stellar classification in one detector and supplies source catalogs to a separate astrometric solver.

Jocelyn: Framework Overview PlateYOLO unifies the conventional pipeline by unifying source extraction and source classification into a single stellar object detection step.

Vera: The framework adapts the YOLOv5 baseline with three structural modifications tailored for micro-scale targets.

Paper summary: Vera: So, to recap the introduction to this Plate-YOLO framework, it’s designed because historical astronomical plates offer a long temporal baseline for time-domain astrophysics, but they get damaged by structural artifacts and degradation.

Jocelyn: The authors present Plate-YOLO as an end-to-end object detection framework that combines source detection and stellar classification into a single detector step, which then supplies source catalogs to a separate astrometric solver.

Subrahmanyan: The core claim is that this unified detector, built on YOLOv5, uses specific modifications—like the Space-toDepth module and the Micro-Scale High-Resolution Detection Head—to specifically handle micro-scale targets found in these plates.

Vera: They use a joint loss function to supervise all three things at once: bounding box regression, objectness, and class prediction. This suggests they are training the network to be very precise across those detection and classification tasks simultaneously.

Jocelyn: The data preparation involves taking digitized plates from China’s historical archives and performing non-linear dynamic range compression using an inverse hyperbolic sine transformation to smooth out intensity distributions.

Subrahmanyan: Then, they partition the images into two hundred fifty-six by two hundred fifty-six pixel sub-image tensors to get the standardized input format for Plate-YOLO, which is necessary before they generate ground truth labels.

Vera: The ground truth labels are generated by running SExtractor with a low detection threshold and cross-matching those candidate detections with Gaia DR3 reference sources to objectively create the training data.

Jocelyn: To refine their results, they introduce a Gaia-guided fine-tuning strategy where they build an archive-specific dataset using plates that already had successful astrometric calibration to generate high quality labeled samples.

Subrahmanyan: This means the network is jointly trained on the standard source dataset and this specialized target dataset, with a regularization term added to stop it from forgetting general representations during this fine-tuning process.

Vera: The overall thesis is that by integrating these components, Plate-YOLO can robustly recover astrometric information from plates that are otherwise substantially degraded by complex structural noise and various artifacts.

Jocelyn: They show this success by demonstrating cross-domain generalization on the UBAI collection, highlighting successful astrometric recovery of plates that previously failed traditional methods.

Conclusion: Vera: Looking at the full title, Plate-YOLO is positioning itself as an end-to-end object detection framework specifically for robust astrometric registration of historical plates. It’s about making the entire pipeline from finding a star to getting its position as accurate as possible on old photographic material.

Jocelyn: The authors are Quanfeng Xu and the rest of the team, and their work points toward a practical solution for dealing with physical degradation in archival data that traditionally required painstaking manual adjustments.

Subrahmanyan: What this means for cosmology is that if we can reliably process these old plates, we get more robust historical constraints on stellar positions, which feeds into our larger models about structure formation and evolution.

Vera: It’s about making the data retrieval process less dependent on archive-specific manual parameter tuning and more reliant on a unified deep learning approach that learns to handle the noise directly.

Jocelyn: So, for someone just listening, it means they are developing a tool that automates the complex task of extracting accurate positions from noisy historical images using modern detection techniques.

Subrahmanyan: It suggests that future work should focus on whether this generalization holds up when applied to entirely new types of plate archives or if these archive-specific adaptations remain necessary for optimal results.

Shanghai Astronomical Observatory, Chinese Academy of Sciences · School of Astronomy and Space Science, University of Chinese Academy of Sciences · Key Lab for Astrophysics, Shanghai · Institute for Frontiers in Astronomy and Astrophysics, Beijing Normal University · School of Physics and Astronomy, Beijing Normal University

astro-ph.IM

Submitted: 2026-10-19

Updated: 2026-10-19

Comments: 14 pages, 6 figures, accepted for publication in The Astronomical Journal

DOI: 10.3847/1538-3881/aebbbe

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: Historical astronomical plates provide a long temporal baseline for time-domain astrophysics, but degradation and structural artifacts can hinder automated astrometric calibration.

Key concepts

PlateYOLO
A framework that combines object detection (finding sources) and stellar classification into one detector. It adapts YOLOv5 with specific modules like SPD and SimAM to handle micro-scale targets on historical plates.
Space-to-Depth (SPD) module
A structural modification added to the network designed to help the model understand the relationship between a source's position on the plate and its actual three-dimensional depth. This helps in accurately locating sources despite optical distortions.
Gaia-guided fine-tuning
A strategy where the model is trained on a set of plates already successfully calibrated using Gaia reference catalogs. This adaptation bridges the gap between general training data and the specific distortions found in historical archives, improving performance.

Terminology

Summary

Historical astronomical plates provide a long temporal baseline for time-domain astrophysics, but degradation and structural artifacts can hinder automated astrometric calibration. p. 0

Where the paper says it

, Shanghai 200030, PR China 2School of Astronomy and Space Science, University of Chinese Academy of Sciences, 1 East Yanqi Lake Rd. , Beijing 100049, PR China 3Key Lab for Astrophysics, Shanghai, 200234, China 4 Institute for Frontiers in Astronomy and Astrophysics, Beijing Normal University, Beijing, 102206, PR China 5School of Physics and Astronomy, Beijing Normal University, Beijing, 100875, PR China ABSTRACT Historical astronomical plates provide a long temporal baseline for time-domain astrophysics, but degradation and structural artifacts can hinder automated astrometric calibration.

PlateYOLO is an end-to-end object detection framework that combines source detection and stellar classification in one detector and supplies source catalogs to a separate astrometric solver. p. 0
Where the paper says it

, Beijing 100049, PR China 3Key Lab for Astrophysics, Shanghai, 200234, China 4 Institute for Frontiers in Astronomy and Astrophysics, Beijing Normal University, Beijing, 102206, PR China 5School of Physics and Astronomy, Beijing Normal University, Beijing, 100875, PR China ABSTRACT Historical astronomical plates provide a long temporal baseline for time-domain astrophysics, but degradation and structural artifacts can hinder automated astrometric calibration. We present PlateYOLO, a framework that combines source detection and stellar classification in one detector and supplies source catalogs to a separate astrometric solver.

Framework Overview

PlateYOLO unifies the conventional pipeline by unifying source extraction and source classification into a single stellar object detection step. The framework adapts the YOLOv5 baseline with three structural modifications tailored for micro-scale targets. p. 4

Where the paper says it

Khanam & M. Hussain 2024) with three structural modifications tailored for micro-scale targets (Figure 2).

These modifications include the Space-toDepth (SPD) module, a Simple Parameter-Free Attention Module (SimAM), and a Micro-Scale High-Resolution Detection Head (P2). p. 4
Where the paper says it

An overview of the Plate-YOLO architecture, highlighting the integration of the Space-to-Depth (SPD) module, SimAM attention mechanisms, and the dedicated Micro-scale Head (P2). This prevalence of small sources motivates the use of higher-resolution features and a regression loss tailored to small bounding boxes.

The joint loss function combines bounding-box regression, objectness, and class prediction to supervise the network. p. 5
Where the paper says it

Joint Loss Function The Plate-YOLO training objective combines bounding-box regression, objectness (foreground vs. background), and class prediction (stellar vs.

Data Preparation and Labeling

The training data come from China’s digitized historical night-time astronomical plates. To convert raw digitized plates into a structured training dataset, two primary operations are performed. p. 2

Where the paper says it

287′′ for DAZ6 . Data Pre-processing and Label Generation To convert raw digitized plates into a structured training dataset, our data preparation comprises two primary operations.

These operations include non-linear dynamic range compression using an inverse hyperbolic sine transformation to mitigate positive skewness in pixel intensity distributions. p. 2
Where the paper says it

To mitigate this issue, we employ a non-linear intensity compression procedure that remaps the pixel values into a dynamic range characteristic of natural images. Pixel values in each single-band patch were clipped at the 1st and 99th percentiles, normalized to [0, 1], and stretched using an inverse hyperbolic sine transformation: Iout = asinh h β Iclip−P1 P99−P1 i asinh(β) , β = 10.

Furthermore, spatial partitioning into 256 × 256 pixel sub-image tensors is used to prepare the data for Plate-YOLO. p. 3
Where the paper says it

Spatial Windowing and Labels Digitized historical astronomical plates routinely exceed 8000 × 8000 pixels, making direct deep learning ingestion computationally intractable. To prepare the data for Plate-YOLO, we implemented an overlapping sliding-window strategy, partitioning the full-field images into 256 × 256 pixel sub-image tensors.

Ground-truth labels are objectively generated by running SExtractor with a low detection threshold and cross-matching candidate detections with Gaia DR3 reference sources. p. 3
Where the paper says it

Candidate detections were cross-matched with Gaia DR3 ( Gaia Collaboration et al. 2023) using reference sources with G ≤ 18 mag 7 , whose coordinates were propagated to each plate epoch using Gaia proper motions.

Model Refinement Strategy

To bridge the domain gap between source domains, a Gaia-guided fine-tuning strategy is introduced. This strategy constructs an archivespecific dataset using plates with successful astrometric calibration to generate high-quality labeled samples. p. 6

Where the paper says it

To bridge this cross-domain gap without requiring manual annotation, we introduce a self-calibration strategy that exploits this partial success (the detailed pipeline is discussed in Section 4. Specifically, we construct an archivespecific dataset using a subset of plates with successful astrometric calibration.

The network is then fine-tuned by jointly training on the source dataset and the target dataset, using a regularization term to prevent catastrophic forgetting of generalized representations. p. 6
Where the paper says it

These high-quality labeled samples from the target domain (Dt) are then used alongside the source dataset (Ds) to fine-tune the Plate-YOLO detector. The regularization term penalizes deviations from the source-model weights, while joint training on Ds and Dt is intended to mitigate forgetting during target-archive adaptation.

Performance Evaluation

The model was evaluated on 150 UBAI plates, where it solved 40 plates initially and recovered a further 106 through fine-tuning, resulting in an overall recovery rate of 146/150 (97.3%). The newly recovered solutions have a median plate-level RMS of 0. p. 9

Where the paper says it

The positional residual is ri = [(∆α ∗ i ) 2 + (∆δi) 2 ] 1/2 , and the plate-level two-dimensional RMS is RMS = vuut 1 N X N i=1 r 2 i , (11) where N is the number of matched sources used in the evaluation. Across these plates, the median RMS is 0.

563'', showing no pronounced outward increase in the median positional residuals. Furthermore, Plate-YOLO achieved 2. p. 9
Where the paper says it

Gaia G magnitude (or all available sources if fewer than 500 were matched). Plate-YOLO achieved a slightly higher median Spearman correlation with G than SExtractor: 0.

53-fold faster source extraction and a 2. p. 0
Where the paper says it

On 150 UBAI plates for which the conventional pipeline had failed, the Chinese-trained model solved 40 plates, and target-archive fine-tuning recovered a further 106, giving an overall recovery rate of 146/150 (97. In the reported hardware configuration, source extraction took 6.

62-fold speedup for the combined extraction and astrometric-solving workflow on the tested hardware. p. 8
Where the paper says it

The checkpoint from the final epoch was used as the Fine-tune Model. The fine-tuned detector combined with brightness ordering and astrometric solving recovered 106 of the 110 plates.

The fine-tuned model reached a peak ratio of about 95% in Gaia-matched stellar detections compared to about 65% for the Base Model. p. 8
Where the paper says it

In each instrumental-magnitude bin, we calculate the ratio of Gaia-matched Plate-YOLO stellar detections to Gaia-confirmed SExtractor sources. The Fine-tuned Model reaches a peak ratio of about 95%, compared with about 65% for the Base Model, and stays above 60% across most of the well-populated bright-source range.

Conclusion and Significance

PlateYOLO integrates source detection and stellar classification for degraded historical photographic plates using SPD feature rearrangement, SimAM attention, a P2 detection head, and NWD regression. The results demonstrate that archive-specific adaptation can recover astrometric information from plates affected by optical distortions, emulsion defects, stellar trailing, and other forms of degradation that hinder conventional source extraction. p. 9

Where the paper says it

635′′), based on a median of 1724 Gaia matches per plate (interquartile range: 602–4253). These residuals are consistent with the stronger image degradation in this challenging sample, including PSF distortions and emulsion defects.

PlateYOLO therefore provides a practical route toward large-scale reprocessing of heterogeneous historical plate archives while reducing the need for archive-specific manual parameter tuning. p. 1
Where the paper says it

PlateYOLO addresses these challenges through small-object detection and archive-specific fine-tuning. In this paper, we propose Plate-YOLO, an object detection framework designed to facilitate the robust astrometric registration of historical photographic plates.

The present performance remains dependent on the available reference catalogs, training-label construction, and the plate populations examined here. p. 8
Where the paper says it

Detection Completeness and Catalog Purity We evaluated detection performance on the 106 newly recovered plates, which were excluded from training and fine-tuning. Using the recovered WCS solutions, we cross-matched the SExtractor and Plate-YOLO catalogs with Gaia DR3 positions propagated to each observation epoch.

References

Bertin E., & Arnouts S. p. 0

Where the paper says it

The initial source detection stage typically relies on threshold-based algorithms like SExtractor ( Bertin, E. & Arnouts, S.

1996, Astron. Astrophys. Suppl. Ser., 117, 393, doi: 10.1051/aas:1996164

Cornu D., Semelin B., Salom´e P., et al. 2025, arXiv e-prints, arXiv:2509. p. 0

Where the paper says it

cn, yuy@shao. cn arXiv:2610.

12082, doi: 10.48550/arXiv. p. 0
Where the paper says it

cn, yuy@shao. cn arXiv:2610.

2509.12082

Enke H., Tuvikene T., Groote D., Edelmann H., & Heber U. 2024, A&A, 687, A165, doi: 10. p. 0

Where the paper says it

Enke et al. 2024) integrated over 70,000 historical plates from multiple German observatories.

1051/0004-6361/202348793

Fluke C. J., & Jacobs C. p. 1

Where the paper says it

Ntampaka et al. Jacobs 2020; F.

2020, Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 10, e1349

Gaia Collaboration Vallenari A., Brown A. G. A., et al. 2023, A&A, 674, A1, doi: 10. p. 1

Where the paper says it

Grishin et al. (2023) successfully introduced YOLO-CL to identify galaxy clusters in large-scale optical surveys.

1051/0004-6361/202243940

Grindlay J., Tang S., Los E., & Servillat M. 2011, Proceedings of the International Astronomical Union, 7, 29–34, doi: 10. p. 0

Where the paper says it

2021) focus on high-precision digitization tailored for Solar System kinematics. In parallel with these international endeavors, a comprehensive effort in China has centralized and digitized approximately 30,000 historical plates at the Shanghai Astronomical Observatory using custom high-precision scanners (M.

1017/S1743921312000166

Grishin K., Mei S., & Ili´c S. 2023, A&A, 677, A101, doi: 10. p. 1

Where the paper says it

Grishin et al. (2023) successfully introduced YOLO-CL to identify galaxy clusters in large-scale optical surveys.

1051/0004-6361/202345976

Hambly N., Irwin M., & MacGillivray H. 2001a, Monthly Notices of the Royal Astronomical Society, 326, 1295, doi: 10. p. 0

Where the paper says it

Draft version October 20, 2026 Typeset using LATEX default style in AASTeX7. 1 Plate-YOLO: An End-to-End Object Detection Framework for Robust Astrometric Registration of Historical Plates Quanfeng Xu , 1, 2 Zhengjun Shang,1 Shiyin Shen , 1, 3 Yong Yu,1, 2 Mingyang Ma,4, 5 Zhenghong Tang,1, 2 Meiting Yang,1 Jing Yang,1 and Jianhai Zhao1 1Shanghai Astronomical Observatory, Chinese Academy of Sciences, 80 Nandan Rd.

1111/j.1365-2966.2001.04661.x

Hambly N., MacGillivray H., Read M., et al. 2001b, Monthly Notices of the Royal Astronomical Society, 326, 1279, doi: 10. p. 0

Where the paper says it

Hambly et al. 2001b,a) digitized southern-hemisphere Schmidt plates to produce sub-arcsecond precision reference catalogs.

1111/j.1365-2966.2001.04660.x

He Y., Wu J., Wang W., Jiang B., & Zhang Y. 2023, Publications of the Astronomical Society of Japan, 75, 1311, doi: 10. p. 1

Where the paper says it

Grishin et al. (2023) successfully introduced YOLO-CL to identify galaxy clusters in large-scale optical surveys.

1093/pasj/psad071

Hu J., Shen L., & Sun G. 2018, in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7132–7141, doi: 10. p. 5

Where the paper says it

Micro-Scale High-Resolution Detection Head The standard YOLOv5 architecture uses a three-tier detection head designed to capture large, medium, and small objects corresponding to the levels of the feature map P5, P4, and P3. However, astronomical targets on historical plates often fall significantly below the threshold of what computer vision algorithms classify as ”small” objects.

1109/CVPR.2018.00745

Khanam R., & Hussain M. p. 4

Where the paper says it

Khanam & M. Hussain 2024) with three structural modifications tailored for micro-scale targets (Figure 2).

2024, What is YOLOv5: A deep look into the internal features of the popular object detector, https://arxiv.org/abs/2407.20892

Lang D., Hogg D. W., Mierle K., Blanton M., & Roweis S. 2010, The Astronomical Journal, 139, 1782, doi: 10. p. 1

Where the paper says it

Lang et al. 2010) for a coarse match, followed by a fine World Coordinate System (WCS) solution via cross-matching with modern high-precision catalogs like Gaia.

1088/0004-6256/139/5/1782

Liu W., Erhan D., Szegedy C., et al. 2016, in ECCV, Vol. p. 1

Where the paper says it

2015), SSD (W. 2016), and YOLO (You Only Look Once) (J.

9905, 21–37, doi: 10.1007/978-3-319-46448-0

Lupton R. H., Gunn J. E., & Szalay A. S. 1999, The Astronomical Journal, 118, 1406, doi: 10. p. 3

Where the paper says it

Lupton et al. 1999): it behaves approximately linearly at low signal levels, thereby preserving the morphology of faint PSF profiles, while approaching a logarithmic response for bright or saturated regions, thus preventing their dominance in the dynamic range.

1086/301004

Ma M., Yuan H., Xiao K., et al. 2025, The Astrophysical Journal Supplement Series, 280, 18, doi: 10. p. 1

Where the paper says it

STAR-YOLO (S. Xu 2025) employed multi-scale attention mechanisms and an Adaptive Focal Spatial-IoU loss function to pinpoint irregular and small-target galaxies accurately.

3847/1538-4365/adec7e

Masias M., Freixenet J., Llad´o X., & Peracaula M. 2012, Monthly Notices of the Royal Astronomical Society, 422, 1674, doi: 10. p. 0

Where the paper says it

Draft version October 20, 2026 Typeset using LATEX default style in AASTeX7. 1 Plate-YOLO: An End-to-End Object Detection Framework for Robust Astrometric Registration of Historical Plates Quanfeng Xu , 1, 2 Zhengjun Shang,1 Shiyin Shen , 1, 3 Yong Yu,1, 2 Mingyang Ma,4, 5 Zhenghong Tang,1, 2 Meiting Yang,1 Jing Yang,1 and Jianhai Zhao1 1Shanghai Astronomical Observatory, Chinese Academy of Sciences, 80 Nandan Rd.

1111/j.1365-2966.20742.x

Ntampaka M., Avestruz C., Boada S., et al. 2021, The Role of Machine Learning in the Next Decade of Cosmology, https://arxiv. p. 1

Where the paper says it

Masias et al. As modern astronomy enters an era of massive data from next-generation surveys, automated deep learning techniques (M.

org/abs/1902.10159

Redmon J., Divvala S., Girshick R., & Farhadi A. p. 1

Where the paper says it

Redmon & A. Farhadi 2018) are capable of performing source detection and classification simultaneously in a single step.

p. 1
Where the paper says it

Redmon & A. Farhadi 2018) are capable of performing source detection and classification simultaneously in a single step.

2016, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 779–788, doi: 10. p. 1
Where the paper says it

2015), SSD (W. 2016), and YOLO (You Only Look Once) (J.

1109/CVPR.2016.91

Redmon J., & Farhadi A. 2018, YOLOv3: An Incremental Improvement, https://arxiv. p. 0

Where the paper says it

cn, yuy@shao. cn arXiv:2610.

org/abs/1804.02767

Ren S., He K., Girshick R., & Sun J. 2015, in Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume 1, NIPS’15, Cambridge, MA, USA, 91–99

Robert V., Desmars J., Lainey V., et al. 2021, Astronomy & Astrophysics, 652, A3, doi: 10. p. 0

Where the paper says it

, Shanghai 200030, PR China 2School of Astronomy and Space Science, University of Chinese Academy of Sciences, 1 East Yanqi Lake Rd. , Beijing 100049, PR China 3Key Lab for Astrophysics, Shanghai, 200234, China 4 Institute for Frontiers in Astronomy and Astrophysics, Beijing Normal University, Beijing, 102206, PR China 5School of Physics and Astronomy, Beijing Normal University, Beijing, 100875, PR China ABSTRACT Historical astronomical plates provide a long temporal baseline for time-domain astrophysics, but degradation and structural artifacts can hinder automated astrometric calibration.

1051/0004-6361/202140472

Shang Z.-J., Yu Y., Wang L.-L., et al. 2024, Research in Astronomy and Astrophysics, 24, 055010, doi: 10. p. 0

Where the paper says it

, Shanghai 200030, PR China 2School of Astronomy and Space Science, University of Chinese Academy of Sciences, 1 East Yanqi Lake Rd. , Beijing 100049, PR China 3Key Lab for Astrophysics, Shanghai, 200234, China 4 Institute for Frontiers in Astronomy and Astrophysics, Beijing Normal University, Beijing, 102206, PR China 5School of Physics and Astronomy, Beijing Normal University, Beijing, 100875, PR China ABSTRACT Historical astronomical plates provide a long temporal baseline for time-domain astrophysics, but degradation and structural artifacts can hinder automated astrometric calibration.

1088/1674-4527/ad339d

Shang Z.-J., Yang M.-T., Wang L.-L., et al. 2026, Publications of the Astronomical Society of the Pacific, 138, 025002, doi: 10. p. 0

Where the paper says it

Draft version October 20, 2026 Typeset using LATEX default style in AASTeX7. 1 Plate-YOLO: An End-to-End Object Detection Framework for Robust Astrometric Registration of Historical Plates Quanfeng Xu , 1, 2 Zhengjun Shang,1 Shiyin Shen , 1, 3 Yong Yu,1, 2 Mingyang Ma,4, 5 Zhenghong Tang,1, 2 Meiting Yang,1 Jing Yang,1 and Jianhai Zhao1 1Shanghai Astronomical Observatory, Chinese Academy of Sciences, 80 Nandan Rd.

1088/1538-3873/ae3cc2

Simcoe R., Grindlay J. p. 0

Where the paper says it

For instance, the Harvard DASCH project (J. Grindlay et al.

E., Los E., et al. 2006, in Applications of Digital Image Processing XXIX, Vol. p. 0
Where the paper says it

Simcoe et al. 2006) to digitize approximately 450,000 plates.

6312, SPIE, 338–349

Sunkara R., & Luo T. 2023, in Machine Learning and Knowledge Discovery in Databases, 443–459

Villaescusa-Navarro F., Angl´es-Alc´azar D., Genel S., et al. 2021, Multifield Cosmology with Artificial Intelligence, https://arxiv. p. 0

Where the paper says it

cn, yuy@shao. cn arXiv:2610.

org/abs/2109.09747

Wang C.-Y., Mark Liao H.-Y., Wu Y.-H., et al. 2020, in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 1571–1580, doi: 10. p. 1

Where the paper says it

Ntampaka et al. Jacobs 2020; F.

1109/CVPRW50498.2020.00203

Xu C., Wang J., Yang W., et al. 2022, ISPRS Journal of Photogrammetry and Remote Sensing, 190, 79, doi: https://doi. p. 5

Where the paper says it

We use the Normalized Wasserstein Distance (NWD; C. 2022) to measure differences in box position and size even for non-overlapping boxes.

org/10.1016/j.isprsjprs.2022.06.002

Xu Q., Shang Z., Shen S., et al. 2026, Research in Astronomy and Astrophysics, https://doi. p. 0

Where the paper says it

, Shanghai 200030, PR China 2School of Astronomy and Space Science, University of Chinese Academy of Sciences, 1 East Yanqi Lake Rd. , Beijing 100049, PR China 3Key Lab for Astrophysics, Shanghai, 200234, China 4 Institute for Frontiers in Astronomy and Astrophysics, Beijing Normal University, Beijing, 102206, PR China 5School of Physics and Astronomy, Beijing Normal University, Beijing, 100875, PR China ABSTRACT Historical astronomical plates provide a long temporal baseline for time-domain astrophysics, but degradation and structural artifacts can hinder automated astrometric calibration.

org/10.1088/1674-4527/ae5f6a

Yang L., Zhang R.-Y., Li L., & Xie X. 2021, in International Conference on Machine Learning

Yang M., Yu Y., Wang L., et al. 2024, Astronomical Techniques and Instruments, 1, 71, doi: 10. p. 1

Where the paper says it

Masias et al. As modern astronomy enters an era of massive data from next-generation surveys, automated deep learning techniques (M.

61977/ati2024005

Yu Y., Zhao J.-H., Tang Z.-H., & Shang Z. p. 0

Where the paper says it

Draft version October 20, 2026 Typeset using LATEX default style in AASTeX7. 1 Plate-YOLO: An End-to-End Object Detection Framework for Robust Astrometric Registration of Historical Plates Quanfeng Xu , 1, 2 Zhengjun Shang,1 Shiyin Shen , 1, 3 Yong Yu,1, 2 Mingyang Ma,4, 5 Zhenghong Tang,1, 2 Meiting Yang,1 Jing Yang,1 and Jianhai Zhao1 1Shanghai Astronomical Observatory, Chinese Academy of Sciences, 80 Nandan Rd.

-J. 2017, Research in Astronomy and Astrophysics, 17, 28, doi: 10. p. 0
Where the paper says it

, Shanghai 200030, PR China 2School of Astronomy and Space Science, University of Chinese Academy of Sciences, 1 East Yanqi Lake Rd. , Beijing 100049, PR China 3Key Lab for Astrophysics, Shanghai, 200234, China 4 Institute for Frontiers in Astronomy and Astrophysics, Beijing Normal University, Beijing, 102206, PR China 5School of Physics and Astronomy, Beijing Normal University, Beijing, 100875, PR China ABSTRACT Historical astronomical plates provide a long temporal baseline for time-domain astrophysics, but degradation and structural artifacts can hinder automated astrometric calibration.

1088/1674-4527/17/3/28

Yuldoshev Q., Shang Z.-J., Turaev S., et al. 2026, Monthly Notices of the Royal Astronomical Society, stag335, doi: 10. p. 0

Where the paper says it

Draft version October 20, 2026 Typeset using LATEX default style in AASTeX7. 1 Plate-YOLO: An End-to-End Object Detection Framework for Robust Astrometric Registration of Historical Plates Quanfeng Xu , 1, 2 Zhengjun Shang,1 Shiyin Shen , 1, 3 Yong Yu,1, 2 Mingyang Ma,4, 5 Zhenghong Tang,1, 2 Meiting Yang,1 Jing Yang,1 and Jianhai Zhao1 1Shanghai Astronomical Observatory, Chinese Academy of Sciences, 80 Nandan Rd.

1093/mnras/stag335

Zhu S.-Q., & Xu Z.-J. 2025, Research in Astronomy and Astrophysics, 25, 125014, doi: 10. p. 0

Where the paper says it

, Shanghai 200030, PR China 2School of Astronomy and Space Science, University of Chinese Academy of Sciences, 1 East Yanqi Lake Rd. , Beijing 100049, PR China 3Key Lab for Astrophysics, Shanghai, 200234, China 4 Institute for Frontiers in Astronomy and Astrophysics, Beijing Normal University, Beijing, 102206, PR China 5School of Physics and Astronomy, Beijing Normal University, Beijing, 100875, PR China ABSTRACT Historical astronomical plates provide a long temporal baseline for time-domain astrophysics, but degradation and structural artifacts can hinder automated astrometric calibration.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. The improved system can perform end-to-end object detection by unifying source extraction and classification into a single step, bypassing multi-step procedural dependencies mentioned in Section 2 of the paper.

  2. The system can utilize a Space-toDepth (SPD) module to perform deterministic rearrangement of spatial information, preserving fine-grained cues from micro-scale celestial targets that are lost in traditional stride-2 max pooling or strided convolutions.

  3. The system can employ a Simple Parameter-Free Attention Module (SimAM) at multi-scale feature fusion nodes to suppress structural irregularities and amplify genuine stellar signals by calculating 3D attention weights based on channel-wise spatial statistics.

  4. The system can utilize a Micro-Scale High-Resolution Detection Head (P2 head) to provide a finer spatial grid for localization, specifically improving sensitivity to sources occupying only a few pixels.

  5. The system can incorporate the Normalized Wasserstein Distance loss (NWD) for bounding box regression, which measures differences in box position and size even for non-overlapping boxes, making the regression loss less sensitive to small pixel-level offsets for tiny targets.

  6. The system can implement a Joint Loss Function that combines bounding-box regression (Lbox NWD), objectness (Lobj using dynamically scaled Focal Loss), and stellar classification (Lcls using Binary Cross-Entropy) to supervise the network effectively.

  7. The system can employ a Gaia-guided fine-tuning strategy by constructing an archivespecific dataset Dt from plates with successful astrometric calibration to generate accurate local reference labels, thereby bridging the cross-domain gap between pre-trained and target archives.

  8. The system can estimate an instrumental brightness proxy from the original, unstretched FITS image using a background-subtracted image-value sum (Equation 9) to rank sources for astrometric solving prior to passing them to Astrometry.net.

  9. The system can achieve faster processing times, requiring only 2.86 s for image preprocessing and 3.75 s for inference per plate, which is over two and a half times faster than the conventional SExtractor pipeline for source extraction alone (6.61 s).

  10. The system can achieve improved astrometric quality on degraded plates, achieving a median plate-level RMS of 0.563 arcseconds after fine-tuning, with no pronounced outward increase in positional residuals across radial bins.

Abstract

Historical astronomical plates provide a long temporal baseline for time-domain astrophysics, but degradation and structural artifacts can hinder automated astrometric calibration. We present Plate-YOLO, a framework that combines source detection and stellar classification in one detector and supplies source catalogs to a separate astrometric solver. The detector adapts YOLOv5 with Space-to-Depth feature rearrangement, SimAM attention, a high-resolution P2 head, and a Normalized Wasserstein Distance loss for small bounding boxes. A Gaia-guided fine-tuning procedure uses initially solved plates to construct labels for a new archive. On 150 UBAI plates for which the conventional pipeline had failed, the Chinese-trained model solved 40 plates, and target-archive fine-tuning recovered a further 106, giving an overall recovery rate of 146/150 (97.3%). In the reported hardware configuration, source extraction took 6.61 s per plate for Plate-YOLO and 16.73 s for SExtractor. These results demonstrate the practical value of adapting source detection to degraded historical plates.

Sources

Related papers