无人机目标检测多深度混合特征域泛化方法研究  

Study on multi-depth mix feature domain generalization method for UAV objection detection

作　　者：王宝亮 姜智 王健 张宝 马振宇 王博航 于海松 WANG Baoliang;JIANG Zhi;WANG Jian;ZHANG Bao;MA Zhenyu;WANG Bohang;YU Haisong(Xi’an Institute of Modern Control Technology,Xi’an 710065,China)

机构地区：[1]西安现代控制技术研究所,西安710065

出　　处：《兵器装备工程学报》2024年第S2期295-302,共8页Journal of Ordnance Equipment Engineering

摘　　要：由于当前无人机目标检测模型与频域等特征的关联能力弱、真实世界场景下知识迁移困难等,导致了模型泛化能力差,提出了一种无人机目标检测的多深度混合特征域泛化方法,强化了不同类型特征间的关联性提高了目标检测模型的泛化能力。提出了针对频域混合特征的混合特征融合方法,可对混合特征数据间的数据关联进行有效强化。为了降低域偏移对模型泛化性的影响,设计了针对多特征域解耦的混合特征多深度跳跃式融合编解码网络。相较于现有方法,可有效处理真实世界未见场景中的无人机目标检测,检测精度有明显提升。In the task of real world unseen scenes objection detection,existing UAV target detection methods have the problem of poor model generalization ability,which is lead by weak correlation ability for frequency domain features and the difficulty of real world scene knowledge transfer.In this paper,we propose a multi-depth hybrid feature domain generalization method for UAV target detection,which can strengthen the correlation between different types of features.Specifically,to strengthen the data correlation between the mixed feature data,we design a hybrid feature fusion method for the mixed feature in frequency domain.To reduce the influence of domain migration on model generalization,we propose a hybrid feature multi-depth skip fusion codec network for multi-feature domain decoupling.Compared with the existing method,our method can effectively deal with the UAV target detection in the unseen scene in the real world.And the detection accuracy has been significantly improved.

关 键 词：无人机目标检测 域泛化 频域特征 混合特征融合 多深度分层编码 

分 类 号：TJ02[兵器科学与技术—兵器发射理论与技术] TP399[自动化与计算机技术—计算机应用技术]