Session
Session VI: FJR Student Competition
Location
Utah State University, Logan, UT
Abstract
We present SIFTER: Star Identification From Transformer-Encoded Representations, a novel neural network-based star identification algorithm for small satellites, trained on an original dataset of synthetic star scenes. SIFTER uses an encoder neural network model to construct sophisticated representations of star patterns by relating centroids of a given star scene. By avoiding the use of manual feature extraction and pattern construction algorithms, SIFTER is more robust to image noise and environmental disturbances than previously proposed classical and neural network-based algorithms. SIFTER generalizes well on real images taken from satellites and observatories, maintaining availability above 97% and error rate below 2% in the presence of noise such as centroid perturbation, false stars, and magnitude error. Furthermore, SIFTER’s underlying model is small, less than half the size of surveyed neural network-based approaches, leading to lightweight storage requirements below 15 MB and an inference time of 12 ms on a desktop computer. These results suggest SIFTER can support star tracking with higher availability and lower resource demands than most previously proposed star identification algorithms, particularly for low-compute small satellites in noisy environments.
SIFTER: Star Identification From Transformer-Encoded Representations
Utah State University, Logan, UT
We present SIFTER: Star Identification From Transformer-Encoded Representations, a novel neural network-based star identification algorithm for small satellites, trained on an original dataset of synthetic star scenes. SIFTER uses an encoder neural network model to construct sophisticated representations of star patterns by relating centroids of a given star scene. By avoiding the use of manual feature extraction and pattern construction algorithms, SIFTER is more robust to image noise and environmental disturbances than previously proposed classical and neural network-based algorithms. SIFTER generalizes well on real images taken from satellites and observatories, maintaining availability above 97% and error rate below 2% in the presence of noise such as centroid perturbation, false stars, and magnitude error. Furthermore, SIFTER’s underlying model is small, less than half the size of surveyed neural network-based approaches, leading to lightweight storage requirements below 15 MB and an inference time of 12 ms on a desktop computer. These results suggest SIFTER can support star tracking with higher availability and lower resource demands than most previously proposed star identification algorithms, particularly for low-compute small satellites in noisy environments.
