Technology and Engineering Center for Space Utilization

Synthetic Aperture Radar (SAR) object detection plays a crucial role in remote sensing applications. However, conventional methods often require high computational and memory costs, limiting their deployment in resource constrained environments. The challenges of SAR imagery such as sparse object distribution, speckle noise, and multi-scale variations make it difficult for existing lightweight detectors to achieve both high accuracy and efficiency. To address this issue, we propose LightKD-SAR, a lightweight SAR object detection framework that combines an efficient network architecture with enhanced instance selection based knowledge distillation. Specifically, we design a lightweight detection network using customized inverted residual modules, and further reduce computational complexity through optimized feature extraction and fusion strategies while maintaining robust detection performance. Additionally, we introduce an improved instance selection mechanism combined with multi-dimensional knowledge transfer, focusing on samples with large prediction discrepancies to enhance learning of ambiguous objects and complex backgrounds in SAR images. Extensive experiments on the large-scale SARDet-100k dataset demonstrate that LightKD-SAR achieves a mAP of 50.92% with only 15.7 GFLOPs and 11.43M parameters. Compared with state-of-the-art methods, the proposed framework demonstrates superior trade-off between detection accuracy and computational efficiency, making it well-suited for practical deployment in real-world SAR-based remote sensing systems.

Accurate and real-time 6-DoF localization is mission-critical for autonomous lunar landing, yet existing approaches remain limited: visual odometry (VO) drifts unboundedly, while map-based absolute localization fails in texture-sparse or low-light terrain. We introduce KANLoc, a monocular localization framework that tightly couples VO with a lightweight but robust absolute pose regressor. At its core is a Kolmogorov–Arnold Network (KAN) that learns the complex mapping from image features to map coordinates, producing sparse but highly reliable global pose anchors. These anchors are fused into a bundle adjustment framework, effectively canceling drift while retaining local motion precision. KANLoc delivers three key advances: (i) a KAN-based pose regressor that achieves high accuracy with remarkable parameter efficiency, (ii) a hybrid VO–absolute localization scheme that yields globally consistent real-time trajectories (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\geq$</tex-math></inline-formula>15 FPS), and (iii) a tailored data augmentation strategy that improves robustness to sensor occlusion. On both realistic synthetic and real lunar landing datasets, KANLoc reduces average translation and rotation error by 32% and 45%, respectively, with per-trajectory gains of up to 45%/48%, outperforming strong baselines.

Image captioning endeavors to generate text descriptions from images using the learned cross-modal generator. Automatically generating descriptions for life ecological experiment images of China’s space station can significantly expedite the comprehension of the intricate semantic details within the experimental data, facilitating the efficient management and accurate retrieval of massive experimental data. However, current image caption approaches trained on general image caption datasets cannot be directly applied to life ecological experiment scenes. In this work, we construct a Chinese caption dataset of life ecological experiment images by collecting and annotating the life ecological experiment image products of China’s space station (LEEoCSSCc). Furthermore, we propose an innovative image captioning approach by exploiting cross-modal prediction and data augmentation (IACPDA), which is designed to harness the raw image to guide the semantic constraints of the generated captions from different perspectives. IACPDA facilitates the transformation of both the original image and the resulting caption into a unified semantic space. It then leverages the predictions derived from the original image to offer constructive guidance for the generated captions. Moreover, we use a semantic-guided image augmentation method based on a diffusion model to generate augmented images that are distinct from the original training images while maintaining the essential semantics of the original images. Consequently, these augmented images can be effectively utilized as soft labels, enabling the learning of a more effective generator. The experiments indicate that our method achieves superior performance compared to other state-of-the-art baselines on the LEEoCSSCc. This paper is the first application of image captioning to the analysis of space science experiment images from China’s space station and further offers vital technical support for the understanding of space science experiments.

Video Object Segmentation (VOS) is essential for the automated quantitative analysis of complex mouse behaviors. However, multi-mouse segmentation remains challenging, as highly similar visual features, rapid motion patterns, and frequent inter-object occlusions often cause tracking drift and identity (ID) switching in existing VOS methods, particularly over long video sequences. To address these challenges, we propose Refinement-based Multi-Mouse Realtime Video Segmentation (RMMRVS), a closed-loop framework that improves temporal stability and spatial accuracy in long-term multi-object tracking. RMMRVS extends the CUTIE architecture by introducing a feature-aware confidence prediction module that continuously assesses tracking reliability and detects potential drift online. When confidence degradation is detected, an adaptive refinement mechanism is activated, employing Mask2Former to perform high-precision spatial resegmentation. The refined masks are then reinjected into the VOS memory, restoring a stable and reliable reference for subsequent tracking. This “pre-dict-evaluate-correct” strategy effectively suppresses error accumulation and ID switching over extended sequences. Extensive experiments on challenging multi-mouse datasets demonstrate that RMMRVS consistently outperforms baselines in segmentation accuracy and temporal consistency, highlighting its effectiveness and scalability for reliable quantitative behavioral analysis.

The control instructions of industrial control systems are prone to threats such as unauthorized access and tampering during transmission and interaction, and access control is a fundamental method to protect data security. Due to the cyber-physical integration and availability constraints in industrial control systems, existing access control methods cannot be directly applied. In this paper, we propose an access control policy for control instructions based on the ciphertext policy attribute-based encryption (CP-ABE) under the availability constraints in industrial control systems. First, we analyze the abnormal behaviors of control instructions in process industrial monitoring systems, and model the attributes associated with field control business and integrate them into CP-ABE to achieve fine-grained access control and avoid non-compliant operations. Second, we adopt a trusted computing mechanism to protect the identity trustworthiness of the transmission node; the confidentiality of the transmitted control instruction is guaranteed by the negotiated symmetric key and the key authorization is realized by the CP-ABE. We further optimize the measuring frequency of the trusted measurement and the deployment policy of the access control method to guarantee business availability. Finally, we conduct formal analysis and experimental validation of the proposed method, and the results show that the proposed access control policy can prevent unauthorized access and non-compliant tampering by industrial control devices and achieve trustworthy delivery of control instructions with controlled computational complexity.