TL;DR
This paper introduces a novel SLIC-DBSCAN algorithm to effectively segment star images, separating the sky region from large-area interferences. This is crucial for accurate attitude determination in spacecraft, as traditional methods struggle with complex visual noise. The algorithm leverages luminance contrast, superpixel segmentation, and density-based clustering to robustly identify and exclude interference, improving the reliability of star trackers.
What Problem Does This Paper Address?
Spacecraft rely on star trackers for precise attitude determination, a critical function for navigation and mission success. Star trackers work by identifying star patterns in images. However, these images are often corrupted by various forms of interference, ranging from background noise to more problematic large-area interferences like those caused by energetic particles, satellite components, or even atmospheric conditions (like thin clouds or light pollution in urban observation scenarios). These interferences can obscure stars, leading to false detections or missed stars, ultimately compromising the accuracy of the attitude determination system and potentially leading to mission failure. Existing methods often struggle with the diverse shapes and intensities of these interferences, especially when they are widespread and irregular.
Methods
The proposed algorithm, named SLIC-DBSCAN, tackles interference suppression in a multi-step process:
- Restricted Luminance-Based Contrast (LC) Transformation: This initial step aims to enhance the contrast between background noise and large-area interferences. Unlike standard LC, a "restricted" version is used, which constrains the saliency calculation to reduce the impact of dark background pixels and better highlight potential interference and star regions.
- Simple Linear Iterative Clustering (SLIC) Segmentation: The saliency map generated from the restricted LC is then segmented into superpixels using the SLIC algorithm. SLIC is chosen for its ability to create perceptually uniform superpixels that adhere well to image boundaries, which is beneficial for segmenting irregularly shaped interference regions. Optimized parameters for SLIC are determined through statistical analysis of boundary recall.
- Feature Extraction and DBSCAN Clustering: Features are extracted from each superpixel, specifically the mean of saliency values and the variance of original gray values. These features are then fed into the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm. DBSCAN is effective here because it can group together superpixels with similar features (likely belonging to the sky region) while identifying and isolating superpixels with dissimilar features (likely interference) as noise. This density-based approach is robust to the varying shapes and distributions of interference.
Datasets
The paper uses real star images for experiments and evaluation. Specific details about publicly available benchmark datasets are not provided. The experiments include:
- Star images with strong interference.
- Star images with moonlight, linear interference, and irregular shapes.
- Star images from real night sky observation experiments, specifically mentioning interference from Starlink constellation satellites.
- Star images from on-orbit satellites, including linear interference, regular annulus interference, and irregular shaped interference caused by moonlight.
Key Insights and Ideas
- The importance of pre-processing for interference: The paper strongly emphasizes that effectively segmenting the sky region before star extraction is a crucial prerequisite for reliable star tracker performance, especially in the presence of significant interference.
- "Restricted" LC for better saliency: Modifying the standard LC transformation by adding a constraint significantly improves the saliency map's ability to highlight relevant regions (stars and interference) while downplaying background noise.
- SLIC's suitability for irregular shapes: SLIC's ability to create superpixels that conform to image boundaries makes it a good choice for segmenting the often irregular shapes of interference regions, unlike simpler grid-based methods.
- DBSCAN's power for outlier detection: DBSCAN's strength lies in its ability to identify clusters of dense data points (sky region) while effectively labeling sparse or differently distributed points (interference) as noise, making it ideal for separating the two.
- Feature engineering for separation: The combination of saliency mean and gray variance as features proves effective in distinguishing between sky and interference regions, with interference exhibiting more divergent characteristics.
Key Results
- The proposed SLIC-DBSCAN algorithm successfully removes large-area interference from star images.
- It effectively extracts star spots from the sky region of real star images, even in the presence of complex interferences.
- Experimental results show that the proposed method improves the ratio of available star spots compared to existing algorithms like Mao's algorithm and NSTS.
- The algorithm demonstrates robustness across various interference types (moonlight, linear, irregular, annulus) and is not sensitive to their shapes.
- The probability of true detection is significantly increased, and the probability of miss detection is lowered when using the proposed method, particularly in scenarios where traditional methods fail due to interference.
- The computation time for a given 1024 × 1024 real star image is reported as 8.74183 seconds.
Limitations and Future Work
- Dynamic interference: The paper acknowledges that the current method is primarily focused on static or slowly changing interference. Future work will involve improving the method to handle large-area interference under dynamic conditions.
- Computational optimization: The authors plan to explore optimizing the proposed algorithm and implementing it on high-performance co-processors, such as the LoongSon chip, for faster processing.
- Extension to other applications: The proposed method is suggested to be extendable to image recognition of satellite-carried optical sensors and real night observation experiments.
Summary updated