Abstract: Abstract: Aiming at the scene of on-orbit resource-constrained non-cooperative target feature extracting, a lightweight detection model SatYOLO improved based on YOLOv8 is proposed to reduce computational complexity of extracting visual features. Firstly, the Starlight module is proposed to replace the YOLOv8 backbone reducing the amount of computation and parameters; Secondly, designed CGRF neck replacing original neck to fuse multi-scale feature attention and improve detection performance; Finally, a lightweight ShareLN head is proposed replacing original head to reduce the amount of computation, accelerate the detection speed and ensure accuracy, while layer normalization is introduced to alleviate performance degradation caused by the lightweight design. Experimental simulation results show that compared with original model, SatYOLO reduces the number of parameters by 40.9% and the amount of computation by 50.0% while the accuracy is basically equivalent; The proposed model effectively achieves the lightweight goal of the on-orbit non-cooperative target feature extracting system.

Key words: feature extraction, lightweight, non-cooperative target, YOLOv8, on-orbit perception

摘要： 摘 要: 针对在轨资源受限的非合作目标识别场景，为了降低视觉特征提取模型的计算复杂度，保证目标检测和局部特征识别的精度和实时性要求，本文提出了一种基于YOLOv8改进的轻量化检测模型SatYOLO。首先，提出Starlight网络模块替换YOLOv8主干以减少计算与参数量；其次，设计CGRF颈部网络替换原模型颈部以融合多尺度特征注意力并提升检测性能；最后，提出一种轻量化检测头ShareLN来替换原检测头，在保证准确性的条件下减少运算量以加快检测速度，引入层归一化方法来缓解轻量化设计导致的性能下降。分析实验仿真结果表明，相较初始模型，SatYOLO参数量减少40.9%，计算量减少50.0%时，识别精度与基准模型基本相当；所提模型可有效实现在轨非合作目标特征提取系统的轻量化目标。

关键词: 特征提取, 轻量化, 非合作目标, YOLOv8, 在轨感知