STar-DETR: A Lightweight Real-Time Detection Transformer for Space Targets in Optical Sensor Systems

Abstract

Optical sensor systems are essential for space target detection. However, previous studies have prioritized detection accuracy over model efficiency, limiting their deployment on resource-constrained sensors. To address this issue, we propose the lightweight space target real-time detection transformer (STar-DETR), which achieves a balance between model efficiency and detection accuracy. First, the improved MobileNetv4 (IMNv4) backbone network is developed to significantly reduce the model’s parameters and computational complexity. Second, group shuffle convolution (GSConv) is incorporated into the efficient hybrid encoder, which reduces convolution parameters while facilitating information exchange between channels. Subsequently, the dynamic depthwise shuffle transformer (DDST) feature fusion module is introduced to emphasize the trajectory formed by space target exposure. Finally, the minimum points distance scylla intersection over union (MPDSIoU) loss function is developed to enhance regression accuracy and expedite model convergence. A space target dataset is constructed, integrating offline and online data augmentation techniques to improve robustness under diverse sensing conditions. The proposed STar-DETR model achieves an AP0.5:0.95 of 89.9%, successfully detecting dim and discontinuous streak space targets. Its parameter count and computational complexity are reduced by 64.8% and 41.8%, respectively, highlighting its lightweight design and providing a valuable reference for space target detection in resource-constrained optical sensors.

Keywords

• space targets
• object detection
• RT-DETR
• lightweight
• data augmentation