Source-linked AI summary
Improving Faint Object Detection for Space Situational Awareness with Variational Autoencoders
Angela Cratere, Luca Ghilardi, Vishnu Reddy, Francesco Dell'Olio, Charalampos S. Kouzinopoulos, Roberto Furfaro
TL;DR
Faint moving-object detection in X-GEO optical SSA is hindered by structured backgrounds and dense stellar fields. The paper combines Tiny-U-Net masking with astro-VAE background inpainting and integrates the preprocessing with shift-and-stack detection, improving reconstruction quality and faint-target detectability while preserving moving targets.
Problem
Optical SSA must detect extremely low-SNR moving objects in X-GEO imagery affected by structured backgrounds, stellar fields, and other contamination.
Method
The pipeline combines lightweight Tiny-U-Net stellar masking with partial-convolution variational reconstruction by astro-VAE before downstream detection.
Results
astro-VAE outperforms PCNN across SSIM, PSNR, and MSE while using approximately half its parameters, and controlled shift-and-stack tests substantially increase stacked SNR.
Takeaways & Limitations
The pipeline provides a data-driven preprocessing strategy for suppressing stellar sources and structured backgrounds while preserving moving targets in optical SSA.
Takeaways & Limitations
Future work must quantify effects on false-alarm rate and track purity and investigate model simplification for embedded deployment.
Abstract
from arXiv · showhide
We present a deep-learning pipeline for enhancing the detection of faint moving objects in optical space situational awareness (SSA) imagery through automated star removal and background reconstruction. Detecting low signal-to-noise ratio (SNR) objects remains extremely challenging in optical observations, particularly in the cislunar (X-GEO) environment, where structured sky backgrounds, dense stellar fields, and scattered moonlight significantly degrade the performance of classical detection algorithms. To address this problem, the proposed pipeline combines a lightweight segmentation network (Tiny-U-Net) to generate stellar masks with a partial-convolution variational autoencoder (astro-VAE), designed to learn the statistical distribution of astronomical backgrounds and perform context-aware inpainting of masked regions. The reconstructed background maps can then be used as a preprocessing step to suppress fixed sources and background inhomogeneities prior to detection. As a proof of concept, the approach is integrated with a shift-and-stack scheme and evaluated on real ground-based telescope observations targeting the X-GEO region. Results demonstrate that the method reconstructs star-free backgrounds with high fidelity, while preserving moving targets and significantly enhancing detectability, thereby providing an effective data-driven preprocessing strategy for faint moving-object detection in optical SSA scenarios.
1 Introduction
Optical SSA in the X-GEO environment must detect extremely faint moving objects amid structured backgrounds and dense stellar fields. The paper addresses this need with learned stellar masking and probabilistic background reconstruction integrated into detection.
- X-GEO optical SSA must detect extremely low-SNR targets affected by sky gradients, stray light, instrumental noise, and dense stellar fields.
- Residual stellar contamination and background inhomogeneities can generate spurious responses in integration-based detection methods.
- Probabilistic generative models for background reconstruction remain largely unexplored in real optical SSA imagery and operational detection pipelines.
- The proposed framework combines Tiny-U-Net stellar masking with astro-VAE background reconstruction and evaluates preprocessing within a shift-and-stack detection scheme.
2 Methods
The method uses Tiny-U-Net to identify stellar regions and a partial-convolution variational autoencoder to reconstruct their surrounding sky background. The resulting background estimate is trained with reconstruction, style, structural, source, and variational objectives.
- 2 Methods: The pipeline processes 256×256 tiles, masks stellar regions, and supplies the masked images to astro-VAE for star-free background reconstruction.
- 2.1 Tiny-U-Net: Tiny-U-Net generates stellar masks with 1.94 M parameters, versus 17.27 M for baseline U-Net, reducing computational complexity and memory requirements.
- 2.2 astro-VAE: astro-VAE formulates stellar regions as missing data and uses partial-convolution layers with a probabilistic latent model to reconstruct structured sky backgrounds from context.
- 2.2 astro-VAE: The latent vector is sampled as z = µ + σϵ, with ϵ ∼N(0, 1), enabling gradient-based optimization of the variational reconstruction model.
- 2.2 astro-VAE: Training combines horizontal and vertical reconstruction, style, SSIM, source, and KL regularization terms with empirically selected weights.
3 Results
On real X-GEO telescope imagery, the proposed models produced effective stellar masks and background reconstructions, while astro-VAE preprocessing improved faint-target visibility in shift-and-stack experiments. The evaluation also compared learned reconstruction with U-Net, PCNN, and conventional Background2D processing.
- 3 Results: The dataset contains 290 bias-, dark-, and flat-field-corrected optical images, with 213 used for training and 77 for testing.
- 3.1 Stellar segmentation: Tiny-U-Net maintains comparable recall and overall accuracy to U-Net despite an 88.8% parameter reduction, so it was selected for pipeline integration.
- 3.2 Background reconstruction: Background2D estimates vary with mesh and filter configuration, whereas astro-VAE uses the same learned model without image-specific parameter selection at inference.
- 3.2 Background reconstruction: astro-VAE outperforms deterministic PCNN across SSIM, PSNR, and MSE while using approximately half the parameters, 33.50M versus 70.13M.
- 3.3 Impact on shift-and-stack detection: The shift-and-stack experiment compares raw stacking, background subtraction, and stacking of inpainted frames for injected targets with SNR values from 1.5 to 5.
4 Discussion
The Tiny-U-Net and astro-VAE pipeline automated stellar masking and context-aware background reconstruction for faint-object detection in X-GEO observations. Integrated with shift-and-stack, it increased stacked SNR while preserving moving targets and suppressing stellar sources, although false-alarm and track-purity impacts remain to be quantified.
- Background reconstruction: The reconstructed backgrounds are compared with manually configured Background2D estimates on the same frame, including their corresponding background-subtracted images.Figure 4 presents conventional estimates using different box_size and filter_size settings alongside the astro-VAE prediction.
- Detection performance: Figure 5 compares raw, standard shift-and-stack, background-subtracted, and star-removed results, with preprocessing improving target contrast while fixed stellar sources are suppressed.Standard stacking accumulates background inhomogeneities and turns stars into streaks, whereas the target remains visible after star removal.
- Detection performance: The controlled shift-and-stack experiment showed that preprocessing substantially increased the stacked SNR of faint injected targets while preserving moving targets and suppressing stellar sources.The experiment used an injected target with SNR = 3 and compared standard, background-subtracted, and star-removed processing.
- Pipeline role: The pipeline combines automated stellar masking with astro-VAE reconstruction to model structured sky backgrounds and remove stellar contamination before downstream detection.The discussion identifies Tiny-U-Net and astro-VAE as the components enabling automated masking and context-aware reconstruction.
- Limitations and future work: Future evaluation must quantify effects on false-alarm rate and track purity in track-before-detect methods, while model compression remains relevant for embedded deployment.The paper also identifies low-latency sensor-level processing as a future deployment goal.