Plate-YOLO: An End-to-End Object Detection Framework for Robust Astrometric Registration of Historical Plates
summary
The gist
Historical astronomical plates provide a long temporal baseline for time-domain astrophysics, but degradation and structural artifacts can hinder automated astrometric calibration.
In short
PlateYOLO is an end-to-end framework that detects and classifies stellar objects from degraded historical astronomical plates. It unifies source extraction and classification into a single step using modified YOLOv5 architecture. The system uses data from Chinese plates, fine-tuned with Gaia data to recover astrometric information, achieving high accuracy in plate registration.
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 used across episodes
This episode discusses
- Plate-YOLO: An End-to-End Object Detection Framework for Robust Astrometric Registration of Historical Plates · Paper Radio
- YOLO-CIANNA: Galaxy detection with deep learning in radio data: II. Winning the SKA SDC2 using a generalized 3D-YOLO network
- What is YOLOv5: A deep look into the internal features of the popular object detector
- The Role of Machine Learning in the Next Decade of Cosmology
- YOLOv3: An Incremental Improvement
- Multifield Cosmology with Artificial Intelligence
The paper
Plate-YOLO: An End-to-End Object Detection Framework for Robust Astrometric Registration of Historical Plates · Read on arXiv
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
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.
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.
More episodes
- 2607.17929-gevolution 2.0: GPU-accelerated relativistic N-body simulations for cosmology
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2509.02657-On the synergetic use of Ariel and JWST for exoplanet atmospheric science
- 2609.10042-Modelling Fourier filtering wavefront sensors for PSD-based methods: the AOPERA tool
- 2511.05639-Performance of the AstroPix Prototype Module for the Barrel Imaging Calorimeter at the ePIC Detector and in Space-Based Payloads
- 2610.19909-RPCA: a Resolved PSF reconstruction method for undersampled images
- 2512.03440-An Analysis of LIGO Glitches Using t-SNE During the First Part of the Fourth LIGO-Virgo-KAGRA Observing Run
- 2608.10284-Lunar Reflective Interferometry
- 2610.16633-Institutional broadening of the Japanese astronomy and astrophysics research community during 1996-2025