Star Image Prediction and Restoration under Dynamic Conditions

Di Liu, Xiyuan Chen, Xiao Liu, Chunfeng Shi2019 · DOI

TL;DR

This paper tackles the challenge of maintaining accurate spacecraft attitude measurements from star sensors during high-speed maneuvers. They propose a novel approach combining gyroscope-assisted star image prediction to handle frame loss and an improved Richardson-Lucy deblurring algorithm, enhanced by an ensemble neural network, to restore smeared images. This dual approach significantly improves the star sensor's robustness in dynamic environments.

What Problem Does This Paper Address?

Spacecraft attitude control relies heavily on star sensors for precise orientation measurements. However, during rapid maneuvers or high-dynamic conditions, star sensors face two critical issues: frame loss (missing image frames) and image smearing (blurring of star points). These problems degrade the accuracy of star centroid extraction, leading to unreliable attitude determination and potentially compromising mission safety. While gyroscopes offer high accuracy for short periods, their drift necessitates periodic recalibration by star sensors. This paper aims to bridge the gap by enabling star sensors to function reliably even when encountering these dynamic challenges.

Methods

The paper introduces a two-pronged approach:

  • Star Image Prediction: To combat frame loss, they leverage gyroscope data. By considering the star sensor's lens distortion, they establish a prediction model that uses the gyroscope's angular rates to estimate the future position of star spots. This allows the system to "fill in the blanks" when frames are lost.
  • Star Image Restoration (Deblurring): For smeared images, they employ a non-blind image deblurring (NBID) technique.
  • PSF Calculation: The Point Spread Function (PSF), which describes the blur, is calculated using the gyroscope's angular velocity information, accounting for different rotation axes (X, Y, Z, and combinations).
  • Improved Richardson-Lucy (RL) Algorithm: The standard RL algorithm is enhanced. The key innovation is an Ensemble Back-Propagation Neural Network (EBPNN) that predicts the optimal number of iterations for the RL algorithm. This prediction is based on the Sum of Magnitude of Fourier Coefficients (SUMFC) of the calculated PSF. This avoids the time-consuming manual trial-and-error typically needed to find the optimal iteration count for RL, improving real-time performance. The EBPNN itself is built using an improved bagging method and a just-in-time learning algorithm for better accuracy.

Datasets

The paper does not explicitly mention using publicly available datasets. Instead, it relies on simulated star images. These simulations are generated based on:

  • A simulated spacecraft trajectory.
  • Parameters of a star sensor (field of view, focal length, pixel size, etc.).
  • The Yale Bright Star Catalogue for selecting guide stars.
  • The ground truth of star images is used for comparison.

Key Insights and Ideas

  • Gyroscope as a Dual-Purpose Tool: The gyroscope is not just for attitude measurement but is crucial for both predicting star positions (to handle frame loss) and calculating the PSF for deblurring.
  • Neural Network for RL Optimization: Using an EBPNN to predict the optimal number of RL iterations is a significant contribution, addressing a major bottleneck in RL-based deblurring for real-time applications.
  • Ensemble Learning for Robustness: The use of an ensemble neural network (EBPNN) rather than a single BP network improves generalization and prediction accuracy, which is vital for handling diverse blurring conditions.
  • Lens Distortion Consideration: Explicitly modeling lens distortion in the star spot prediction adds a layer of realism and accuracy to the prediction model.

Key Results

  • Star Image Prediction Accuracy: The proposed prediction method achieved a maximum error of 0.89 pixels horizontally and 0.50 pixels vertically for star spot positions within 15 seconds of prediction.
  • Deblurring Performance: The deblurring method significantly suppressed smearing and improved the gray value and distribution of star spots, making them more detectable.
  • Attitude Estimation Improvement: The restored star images allowed for attitude estimation at much higher angular velocities (e.g., up to 75°/s for X-axis rotation under Gaussian noise) compared to blurred images (where estimation often failed beyond 25°/s). Both the proposed method and an iterative RL method showed similar improvements in attitude estimation range.
  • Real-time Performance: The proposed method demonstrated significantly better real-time performance than the iterative RL algorithm, especially under Gaussian noise, due to the EBPNN's rapid iteration count prediction.

Limitations and Future Work

  • Simulation-Based Validation: While the results are promising, they are primarily based on simulated data. Further validation with real-world flight data would be beneficial.
  • Computational Complexity: Although the EBPNN improves real-time performance, the overall process still involves complex calculations for PSF estimation and neural network inference.
  • Future Work: The paper implicitly suggests that further optimization of the EBPNN architecture and training process could lead to even better performance. Exploring more advanced deblurring techniques or hybrid approaches could also be avenues for future research.

Summary updated

BibTeX
@misc{liu2019star,
  author = {Di Liu and Xiyuan Chen and Xiao Liu and Chunfeng Shi},
  title = {{Star Image Prediction and Restoration under Dynamic Conditions}},
  year = {2019},
  doi = {10.3390/s19081890},
}
Figure