空间变化场景下卫星部组件域适应识别研究

针对小样本限制下卫星部组件识别域适应困难的问题，提出一种变化场景下自适应迁移的目标识别算法。卫星部组件的识别模型框架为YOLO，迁移算法包括3个策略：基于特征关联性的样本加权策略，基于模型的参数自适应策略和最优特征变换自适应迁移策略。基于以上策略，YOLO模型建立域特征空间的相似性，选择性地迁移源域知识，同时在适应过程中通过调整策略边界学习不变特征表示，以此来加强模型的自适应迁移能力。迁移实验中，分别验证了3个策略的迁移能力，有效提升YOLO模型在复杂多变空间环境下对卫星部组件的稳定识别。实验结果表明：基于特征关联性学习到的权重要优于随机初始化权重，参数自适应迁移显著提升了目标域的测试精度，最优特征变换显著提升模型的泛化性能力。

Research on adaptive domain detection of satellite component under space variable environment

To solve the few-shot satellite components domain detection, this paper proposes an object detection algorithm for adaptive migration in variable scenarios. The model is based on YOLO, the improvements include three parts: a sample weighting strategy based on feature relevance, and a model-based parameter adaptive strategy and an optimal feature transformation adaptive migration strategy. Based on the above strategies, YOLO builds the similarity of the domain feature space, selectively migrates the source domain knowledge, and adjusts the boundary of the strategy to learn the invariant feature representation during the adaptation process to enhance the adaptive migration ability of the model. In the migration experiment, the migration ability of the three strategies is verified respectively, which effectively improves the stable detection of YOLO-based satellite components detection in the complex space environment. The experimental results show that the weight importance based on feature association learning is better than the random initial weight, the parameter adaptive transfer significantly improves the testing accuracy of the target domain, and the optimal feature transformation significantly improves the generalization ability of the model.