运用LPE-YOLO算法的海面遥感图像多尺度目标检测方法OA
Multi-scale target detection method for sea-surface remote sensing images using LPE-YOLO algorithm
针对海面遥感图像中舰船目标尺度差异显著、具有大长宽比、任意方向排列以及背景噪声干扰强导致的检测精度低、特征提取不充分等问题,提出一种基于YOLOv8算法改进的多尺度目标检测(LPE-YOLO)算法.在主干网络中引入线性可变卷积(LDConv)模块,增强算法对几何形变目标的特征拟合能力;在特征融合网络中设计并集成C2f-PEMA模块,利用部分卷积(PConv)模块削减计算冗余,并嵌入高效多尺度注意力(EMA)机制,抑制海面杂波干扰,强化对多尺度目标的感知能力.仿真结果表明,在处理HRSC2016与DOTA海面遥感图像混合数据集时,LPE-YOLO算法的目标检测精度达到91.8%,相比基线算法、YOLOv5n算法和YOLO11n算法均有提升.
In order to address the problems such as low detection accuracy and insufficient feature extraction caused by ship targets'significant size-scale differences,large aspect ratios,arbitrary directional arrangements,as well as strong background noise interference in sea-surface remote sensing images,an improved multi-scale target detection algorithm(LPE-YOLO)based on YOLOv8(You Only Look Once version 8)algorithm is proposed.Alin-ear deformable convolution(LDConv)module is introduced into the backbone network to enhance the algorithm's fea-ture fitting ability for geometrically deformed targets.AC2f-PEMAmodule is designed and integrated into the feature fusion network,where partial convolution(PConv)module is utilized to reduce computational redundancy and an effi-cient multi-scale attention(EMA)mechanism is embedded to suppress sea-surface clutter interference and enhance the perception capability of multi-scale targets.Simulation results show that,for the mixed dataset of HRSC2016 and DOTA sea-surface remote sensing images,the LPE-YOLO algorithm achieves a target detection accuracy of 91.8%,which is an improvement compared to baseline algorithm,YOLOv5n,and YOLO11n.
陈可仁;朱旭芳;赵傅彦
海军工程大学电子工程学院,武汉 430030海军工程大学电子工程学院,武汉 430030海军工程大学电子工程学院,武汉 430030
信息技术与安全科学
海面遥感图像目标检测YOLOv8算法LDConv模块C2f-PEMA模块部分卷积模块EMA机制
sea-surface remote sensing imagesobject detectionYOLOv8 algorithmLDConv moduleC2f-PE-MA modulePConv moduleEMA mechanism
《空天预警研究学报》 2026 (2)
105-109,136,6
评论