Source-linked AI summary
Astrometry.net: Blind astrometric calibration of arbitrary astronomical images
Dustin Lang, David W. Hogg, Keir Mierle, Michael Blanton, Sam Roweis
TL;DR
Astronomical image collections need reliable calibration and metadata to become searchable and usable across heterogeneous archives. Astrometry.net uses geometric hashing to infer astrometric calibration directly from image pixels, achieving above 99.9% recognition with no false positives while supporting very small images. Its scope is constrained by image conformality assumptions and by catalog or image features that can produce false matches.
Problem
Large heterogeneous image collections lack fast, robust, automated calibration and detailed calibration metadata needed for effective use and verification.
Method
Astrometry.net detects stars, hashes four-star subsets into geometric features, and searches precomputed indices to generate and verify astrometric-calibration hypotheses.
Results
Above 99.9% recognition with no false positives was achieved for high-quality images, and specialized indices recognized tiny Hubble images containing only a few stars.
Takeaways & Limitations
The system can bring otherwise hidden astronomical data to light and generate consistent astrometric metadata for images from varied sources.
Takeaways & Limitations
The geometric feature generally requires conformal images, while linear source features can produce false matches against linear catalog flaws.
Abstract
from arXiv · showhide
We have built a reliable and robust system that takes as input an astronomical image, and returns as output the pointing, scale, and orientation of that image (the astrometric calibration or WCS information). The system requires no first guess, and works with the information in the image pixels alone; that is, the problem is a generalization of the "lost in space" problem in which nothing--not even the image scale--is known. After robust source detection is performed in the input image, asterisms (sets of four or five stars) are geometrically hashed and compared to pre-indexed hashes to generate hypotheses about the astrometric calibration. A hypothesis is only accepted as true if it passes a Bayesian decision theory test against a background hypothesis. With indices built from the USNO-B Catalog and designed for uniformity of coverage and redundancy, the success rate is 99.9% for contemporary near-ultraviolet and visual imaging survey data, with no false positives. The failure rate is consistent with the incompleteness of the USNO-B Catalog; augmentation with indices built from the 2MASS Catalog brings the completeness to 100% with no false positives. We are using this system to generate consistent and standards-compliant meta-data for digital and digitized imaging from plate repositories, automated observatories, individual scientific investigators, and hobbyists. This is the first step in a program of making it possible to trust calibration meta-data for astronomical data of arbitrary provenance.
1. Introduction
Astronomical data remain difficult to share and use because heterogeneous image collections require automated calibration, vetting, organization, search, and retrieval. Astrometry.net addresses this need by producing astrometric metadata from image pixels without a first guess, enabling otherwise lost imagery to be used.
- Motivation: Large astronomical image collections are underused because sharing data and using others’ shared data remain difficult.The latter two problems are identified as technological challenges.
- Motivation: Fast, robust, automated tools are needed to calibrate, vet, organize, search, and retrieve large heterogeneous image collections.
- Motivation: Virtual Observatory protocols support distributed image search, but publishers must implement standards correctly and detailed calibration processes are not communicated.Users may need to consult papers or contact publishers to determine how raw frames were reduced.
- Contribution: Astrometry.net returns an image’s pointing, scale, and orientation from image pixels alone, without requiring a first guess.These outputs constitute the image’s astrometric calibration or WCS metadata.
- Results: Above 99.9% success with no false positives was achieved for contemporary near-ultraviolet and visual imaging survey data.
- Approach: The system uses geometric hashing to generate calibration hypotheses from sparse star matches, followed by detailed verification.This two-stage structure is shared with related geometric-hashing approaches.
2. Methods
Astrometry.net detects stars, hashes four-star quads, searches indexed codes for calibration hypotheses, and verifies candidates with Bayesian model comparison.
- Hypothesis generation: The search generates candidate alignments by matching query hash codes against a precomputed index, then converts matches into hypothesized position, scale, and orientation.The search heuristic addresses the impracticality of exhaustive exploration of the continuous four-dimensional calibration space.
- Source detection: The system detects compact sources, extracts sub-pixel star positions, and tolerates some missing stars or contaminants in the detection list.Detection includes background flattening, peak identification, cleaning, and Gaussian-based centroiding.
- Hashing of asterisms: Each four-star quad is represented by a geometric hash derived from two stars defining coordinates and two stars’ relative positions.The resulting 4-vector is invariant to translation, rotation, and scaling, enabling comparison across image coordinate systems.
- Hashing of asterisms: Quads are preferred over triangles because typical positional noise makes triangles insufficiently distinctive for reasonable geometric-hashing performance.Noise broadens matching regions in code space, increasing the need for distinctive higher-dimensional codes.
- Bayesian verification: Verification evaluates query stars under foreground and background models, multiplying their likelihood ratios until a threshold is exceeded or the stars are exhausted.The system uses a conservative prior of 10^-6 and strongly penalizes false positives in its utility structure.
3.1. Blind astrometric calibration of the Sloan Digital Sky Survey
The system blind-calibrates SDSS images using indexed geometric asterisms, achieving over 99.97% recognition and 100% after augmenting the reference catalog and correcting source extraction issues.
- Excellent images: The system recognized over 99.97% of excellent SDSS images using only detected-star positions and limited scale information.The experiment discarded astrometric metadata and used quads from the brightest 50 stars.
- Excellent images: Adding a 2MASS-based index recognized the remaining 51 of 61 USNO-B failures, while reprocessing saturated-star cases produced 100% recognition with no false positives.All 182,221 excellent images were ultimately recognized correctly.
- Varying bandpass: Recognition time rises with bandpass mismatch, although the asymptotic recognition fraction remains essentially unchanged after examining the brightest 50 stars.The performance loss is largely attributed to defects in the reference catalog.
- Varying image quality: Bad-quality images show a significant performance drop, but the system still recognizes over 99% of them.Excellent, good, and acceptable images perform almost identically.
- Varying angular size: Recognition degrades slowly down to 8 × 8 arcmin2 and then sharply, reaching zero below 2.8 × 2.8 arcmin2 for this index.The threshold follows from the index requiring stellar separations of at least 4 arcmin to form usable quads.
- Index design: Reducing index density from 16 to 9 quads per HEALPix cell preserves recognition while reducing CPU time to two-thirds.Further density reductions significantly reduce recognition and can increase total CPU time.
- Index design: Quad features provide a computational balance between the false matches of triangles and the larger search cost of triangles and quints.The comparison examined recognition after the brightest 50 stars.
3.2. Blind astrometric calibration of Galaxy Evolution Explorer data
The system was tested on large near-UV GALEX source catalogs using multiple indices spanning the relevant angular scales. Recognition was similar to excellent SDSS r-band performance, but processing was slower.
- Data and setup: The GALEX experiment used near-UV catalogs from 28,182 images with circular fields about 1.2 degrees in diameter.The catalogs were produced by source extraction from GALEX images.
- Results: GALEX recognition was quite similar to excellent-quality SDSS r-band recognition despite the near-UV bandpass being far from the index bandpass.The larger angular size appears to compensate for the bandpass mismatch.
- Results: GALEX recognition was significantly slower because the experiment used a wide angular-scale range and several indices instead of one scale-tuned index.The system was told that images were between 1 and 2 degrees wide and could build quads from the first 100 sources.
3.3. Blind astrometric calibration of Hubble Space Telescope data
Astrometry.net recognized Hubble ACS images using a custom index built for the AEGIS footprint and small quad scales. All 191 inputs were recognized quickly, despite the index’s smallest effective quad being larger than its minimum configured size.
- Index construction: The custom index covered a 2 × 2 degree AEGIS region using about 57,000 SDSS sources and just over 100,000 quads spanning 0.5–2 arcmin.The indexed sources had r-band brightnesses between 15 and 22.2 mag, with HEALPix cells of 0.5 arcmin.
- Recognition results: 100% of the 191 Hubble ACS input images were successfully recognized.The inputs were level 2 images with cosmic rays removed, supplied as 600 × 600-pixel JPEG previews.
- Recognition results: Recognition required an average of 0.3 seconds of CPU time per image, excluding source extraction.
- Recognition results: The recognized images had 64 unique footprints because many footprints were observed through multiple bandpass filters.
- Recognition results: Although indexed quads ranged from 0.5 to 2 arcmin, the smallest quad used to recognize a field was about 0.9 arcmin.
3.4. Blind astrometric calibration of other imagery
Astrometry.net was applied beyond survey images to archival plates, reprocessed survey data, amateur and consumer-camera images, and videos. These applications address imagery whose astrometric metadata is missing or incorrect.
- Photographic archives: The DASCH project uses Astrometry.net to create initial astrometric calibrations for digitized Harvard photographic plates.The archive contains over 500,000 plates exposed between 1880 and 1985, whose original calibrations were handwritten in log books.
- Reprocessed survey data: DeepSky had successfully processed over 14 million Palomar-QUEST and Nearby Supernova Factory images with Astrometry.net.Many of these images had incorrect astrometric metadata.
- Other imagery: The system also successfully calibrated amateur telescope shots, consumer DSLR photographs, and videos posted to YouTube.These images span different exposure properties, optical, infrared, and ultraviolet bands, and can include substantial projection distortions.
3.5. False positives
False positives arose when nonstellar linear features in input images aligned with linear flaws in the USNO-B reference catalog. The figures illustrate how both types of artifact can produce matching lines of detected sources.
- Failure mode: Images that violate the model assumptions can produce false positives at rates much higher than the conservative analysis predicts.
- Input artifacts: Linear features such as an International Space Station streak can be detected as lines of false sources.
- Reference-catalog artifacts: Text on a USNO-B photographic plate can likewise generate many false catalog sources that align with the input artifact.The resulting alignment can be accepted as a false match.
4. Discussion
The discussion presents Astrometry.net as a blind, geometric-hashing calibration system that can restore trustworthy metadata across heterogeneous astronomical imagery. Performance depends on matching indices and image properties, while specialized indices extend recognition to smaller fields and different scales.
- System and contribution: Astrometry.net determines image pointing, orientation, and plate scale without prior information beyond the image pixels.Indexed asterisms generate hypotheses, which are then quantitatively verified.
- Scientific utility: The system can verify or restore astrometric metadata for images of unknown provenance, supporting consistent calibration across heterogeneous collections.The paper connects trusted metadata with large-scale statistical studies enabled by the Virtual Observatory.
- Performance factors: Performance is best when the indexed asterisms match the target images; bandpass differences cause a small performance drop across near-infrared to near-ultraviolet data.Multiple indices could address this if suitable reference catalogs were available.
- Scale dependence: Scale coverage uses multiple indices tuned to narrow ranges, but performance drops sharply outside an index’s tuned range.The main index worked well for 13 × 9 arcmin2 SDSS images and sub-images down to 8 × 8 arcmin2, but dropped seriously at 7 × 7 arcmin2.
- Geometric features: Quad-based indices balance feature distinctiveness against the probability of finding all stars, while triangle or quintuple indices may suit larger or smaller images.
- Specialized indexing: Specialized indices covering tiny sky regions enabled recognition of small HST images containing only a few cataloged stars.
- Scope and limitations: The geometric feature formally requires conformal images, although near-conformal regions can permit recognition despite shear or optical distortion.Non-isotropic scaling and shear generally fall outside the feature’s exact invariance.
- Performance: Over 99.9% recognition with no false positives was achieved for high-quality images.The experiments covered SDSS, GALEX, and HST images.