XAMI -- A Benchmark Dataset for Artefact Detection in XMM-Newton Optical Images
This paper introduces XAMI, a benchmark dataset of 1,000 hand-annotated XMM-Newton Optical Monitoring images, and demonstrates a hybrid CNN-transformer instance segmentation method for the automated detection and masking of light-reflection artefacts in astronomical observations.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine the universe as a giant, glittering ballroom where stars and galaxies are the dancers. Astronomers are the photographers trying to capture perfect portraits of this dance. But sometimes, the camera lens gets smudged, or a stray light bounces off the glass, creating weird ghostly shapes, streaks, or rings on the photo. In the world of astronomy, these unwanted smudges are called artefacts. They are like annoying scribbles on a masterpiece that can trick the photographer into thinking a ghost is dancing when it's just a reflection.
For a long time, trying to clean up these photos was like trying to find a specific needle in a haystack using only a metal detector that was set to "general metal." The old methods relied on rigid rules about how light should behave, but space is messy, and those rules often failed to catch the tricky, weird shapes of the smudges.
Enter the XAMI project (which stands for XMM-Newton optical Artefact Mapping for astronomical Instance segmentation). The team behind this study decided to teach a computer to be a super-smart photo editor. They didn't just teach it to look for "bad things"; they taught it to recognize specific types of cosmic smudges, like Read-Out-Streaks (which look like long lines from the camera's shutter), Smoke Rings (ghostly circles from light bouncing inside the detector), and Star Loops (elongated streaks from bright stars).
To do this, they created a special training school called the XAMI-Dataset. They gathered 1,000 real photos from the XMM-Newton space telescope and, by hand, drew 7,021 precise outlines around the bad smudges. It's like having a teacher show a student 7,000 examples of "this is a scratch, this is a smudge, this is a ring" so the student learns to spot them instantly.
The secret sauce of their method is a "hybrid" approach. Imagine trying to catch a fast-moving ball. You need two things: a quick reflex to spot where the ball is (a CNN or Convolutional Neural Network, specifically a model called YOLOv8), and a steady hand to trace the exact shape of the ball without cutting off its edges (a Transformer model called SAM, or Segment Anything Model).
The paper shows that by combining these two, the computer can first draw a box around the artefact and then use the advanced model to trace the perfect, smooth outline around it. When they tested this on a new set of photos it hadn't seen before, the results were promising but not perfect. The system correctly identified about 84.3% of the artefacts it found (Precision) and managed to catch about 72.1% of all the artefacts that were actually there (Recall).
However, the paper is careful to note that this isn't a magic wand that solves everything yet. The "heavy" part of the system (the SAM model) takes about 70–80 milliseconds per image just to do the heavy lifting, making the whole process take around 100 milliseconds per image. While that's fast enough for processing huge piles of data (like hundreds of thousands of images), it's not quite fast enough for "instant" real-time processing where you need an answer in a split second.
The authors also point out that space photos are tricky because the brightness and exposure times vary wildly. Sometimes it's hard to tell if a faint smudge is a real artefact or just a shadow, which makes it difficult to set a perfect rule for the computer. The "Other" category of artefacts was particularly hard to catch, with the system only finding 33.3% of them, likely because there were so few examples of them in the training school.
In short, the paper suggests that this new hybrid method is a solid, reproducible starting point—a "baseline"—for cleaning up space photos. It proves that mixing old-school object detection with new-school AI segmentation works better than using just one or the other. But the authors are clear: this is a work in progress. They plan to feed the system more data from different space missions and hope that astronomers will help them fine-tune the rules so the computer can become an even sharper editor for the universe's ballroom photos.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.