GPU Accelerated Image Quality Assessment-Based Software for Transient Detection

Xianglong Li, Katarzyna Adamek, Wesley ArmourarXiv 2025 · DOI · arXiv

Abstract

Fast imaging localises celestial transients using source finders in the image domain. The need for high computational throughput in this process is driven by next-generation telescopes such as Square Kilometre Array (SKA), which, upon completion, will be the world's largest aperture synthesis radio telescope. It will collect data at unprecedented velocity and volume. Due to the vast amounts of data the SKA will produce, current source finders based on source extraction may be inefficient in a wide-field search. In this paper, we focus on the software development of GPU-accelerated transient finders based on Image Quality Assessment (IQA) methods -- Low-Information Similarity Index (LISI) and augmented LISI (augLISI). We accelerate the algorithms using GPUs, achieving kernel time of approximately 0.1 milliseconds for transient finding in $2048 \times 2048$ images.

Introduction

In radio astronomy, "transients" are objects that change brightness or appear and disappear over short timescales. Detecting these events is critical for understanding dynamic cosmic phenomena, yet it remains a "needle in a haystack" problem. Traditional methods require comparing wide-field images taken at different times to identify movements or flickers.

The primary challenge lies in data volume. Next-generation telescopes like the Square Kilometre Array (SKA) will produce data at an unprecedented rate. Standard "source extraction" methods—identifying every object in an image and comparing catalogs—are computationally expensive and often too slow for real-time, wide-field searches. This research proposes replacing heavy extraction algorithms with leaner, faster similarity indexes capable of flagging changes instantly.

Methodology

The authors adapt two Image Quality Assessment (IQA) metrics for transient detection:

  1. LISI (Low-Information Similarity Index): Highly sensitive to intensity changes, ideal for spotting faint point sources in low-noise environments.
  2. augLISI (augmented LISI): A robust version designed to handle extended sources (like nebulae) and higher background noise.

To achieve "SKA speed," custom CUDA kernels (split_lisi and split_auglisi) were developed for NVIDIA GPUs. The approach involves:

  • Tiling: Dividing large images into smaller blocks (e.g., $64 \times 64$ pixels).
  • Parallel Reduction: Utilizing GPU shared memory to sum pixel values across tiles simultaneously.
  • Aggregation: A mathematical framework aggregates local tile scores into a single global similarity value, ensuring that parallel processing on small tiles yields the same result as processing the entire image at once.

Experimental Setup

The study utilized simulated datasets to test the algorithms under controlled conditions:

  • GLEAM Catalogue: Real-world source information from the GaLactic and Extragalactic All-sky MWA survey.
  • OSKAR Simulator: Simulates SKA1-LOW telescope observations.
  • MS1 & MS2 Datasets: MS1 contains 33 sources; MS2 contains 28 sources (5 sources were removed to simulate disappearing transients).

Performance and Localization Results

The core innovation is treating transient detection as a similarity problem rather than an object detection task. If the similarity index for a specific tile drops significantly, a transient is flagged.

Computational Performance

On an NVIDIA H100 GPU, the kernels achieved median execution times of 0.101 ms for LISI and 0.091 ms for augLISI on $2048 \times 2048$ images. As shown in the performance scaling plot, the GPU implementation maintains high efficiency even as image sizes scale up toward $8000 \times 8000$ pixels, demonstrating the scalability required for future observatories.

Execution time vs. image size for different tile configurations on an NVIDIA GPU.

Detection and Localization

There is a clear trade-off between sensitivity and robustness. While LISI is extremely sensitive, it often picks up artifacts from the image cleaning process (CLEAN algorithm), leading to false positives. In contrast, augLISI is more robust, ignoring noise to focus on significant source changes. The augLISI heatmap illustrates successful localization of all five "missing" sources in the simulated data with minimal false positives.

Heatmap showing successful localization of five transient sources using the augLISI metric.

Conclusion

This GPU-accelerated IQA approach provides a scalable solution for real-time transient detection in the SKA era. While LISI’s sensitivity can lead to false positives from processing artifacts, augLISI offers a more reliable alternative for complex fields. Future work will focus on integrating these tools into automated pipelines for transient classification, distinguishing between various celestial events like pulsars and supernovae.

Summary updated

Abstract

Fast imaging localises celestial transients using source finders in the image domain. The need for high computational throughput in this process is driven by next-generation telescopes such as Square Kilometre Array (SKA), which, upon completion, will be the world's largest aperture synthesis radio telescope. It will collect data at unprecedented velocity and volume. Due to the vast amounts of data the SKA will produce, current source finders based on source extraction may be inefficient in a wide-field search. In this paper, we focus on the software development of GPU-accelerated transient finders based on Image Quality Assessment (IQA) methods -- Low-Information Similarity Index (LISI) and augmented LISI (augLISI). We accelerate the algorithms using GPUs, achieving kernel time of approximately 0.1 milliseconds for transient finding in 2048X2048 images.
BibTeX
@misc{li2025gpu,
  author = {Xianglong Li and Katarzyna Adamek and Wesley Armour},
  title = {{GPU Accelerated Image Quality Assessment-Based Software for Transient Detection}},
  year = {2025},
  eprint = {2501.10653},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.10653},
  howpublished = {arXiv},
}
Figure