STDSD: A Spacecraft Target Detection Framework Considering Similarity and Diversity

Benefiting from the development of deep learning and computer vision, object detection performance has witnessed remarkable improvements in various domains recently. Nonetheless, several challenges persist for spacecraft detection, including challenging space environment, domain gaps, and real-time application. This paper proposes a Spacecraft Target Detection framework considering Similarity and Diversity strategy (STDSD), specifically designed to adapt and excel in the special requirements of space object detection. The STDSD framework integrates multiple feature extraction networks (FNets) and overlapping vectors prediction networks (ONets). Initially, the original images are cropped into several segments, which are processed through the FNets to extract critical features. These features are subsequently input to the ONets, where adversarial learning and random dropout techniques are implemented to enhance network diversity. This strategy of incorporating both similarity and diversity significantly enhances the precision and robustness of target detection, especially in bridging domain gaps. STDSD was evaluated against the existing detectors using the SPEED, SPEED+, SwissCube, Web-scraped, and Hardware-in-loop datasets. Specifically, STDSD achieved an F-score of 0.9989 on the SPEED dataset, 0.9876 on the “Lightbox” subset, and 0.9523 on the “Sunlamp” subset of the SPEED+ dataset. Additionally, on the SwissCube, Web-scraped, and Hardware-in-loop datasets, the F-scores were 0.9805, 0.8848, and 0.9879, respectively. Furthermore, STDSD exhibited a remarkable capability of over 10 frames per second (FPS) @GPU on a laptop. These results demonstrate that STDSD achieves high prediction accuracy and processing speed, underscoring the robustness capabilities across different datasets and the practicality for real-time applications in space environment. Note to Practitioners—This paper was motivated by the lack of robustness and accuracy in spacecraft object detection, which is also applicable to target detection in other fields. Existing approaches typically employ end-to-end networks that directly output target positions without considering the complex backgrounds and the requirements of real-time and reliability for space missions. In this paper, a spacecraft target detection framework considering similarity and diversity strategy is proposed. It utilizes the uncertainty and randomness outputted by neural networks, where different networks exhibit diversity in suspicious regions and similarity in target regions. By combining the results from multiple networks, the method leads to improved prediction accuracy. In this paper, we analyze the impact of parameters such as dataset distribution, thresholds, and the number of networks on the accuracy of object detection, and compare it with state-of-the-art methods. Experimental results demonstrate that the proposed method has better performance and the detection accuracy and proceeding speed might be balanced by adjusting different network configurations.