面向无人机载平台的轻量级小目标检测算法  

Lightweight Small Object Detection Algorithm for UAV Platforms

作　　者：党兰学[1] 李赞 苗长伟 崔金华 赵雅靓[4] DANG Lanxue;LI Zan;MIAO Changwei;CUI Jinhua;ZHAO Yaliang(Henan Key Laboratory of Big Data Analysis and Processing,Henan University,Henan Kaifeng 475004,China;China Nuclear Industry Survey Design&Research Co.,Ltd,Zhengzhou 475000,China;School of Computer Science and Technology,Huazhong University of Science and technology,Wuhan 430074,China;School of Computer and Information Engineering,Henan University,Henan Kaifeng 475004,China)

机构地区：[1]河南大学河南省大数据分析与处理重点实验室,河南开封475004 [2]中核勘察设计研究有限公司,郑州475000 [3]华中科技大学计算机与信息工程学院,武汉430074 [4]河南大学计算机与信息工程学院,河南开封475004

出　　处：《河南大学学报(自然科学版)》2025年第1期1-11,共11页Journal of Henan University:Natural Science

基　　金：河南省高校科技创新团队支持计划(24IRTSTHN021);河南省研究生教育改革与质量提升工程项目(YJS2024JD30);河南省研究生教育改革与质量提升工程项目(YJS2023JD28)

摘　　要：无人机载平台中的目标检测在军事和民用领域具有重要的应用价值.然而,现有的检测方法通常侧重于多尺度目标检测,缺乏对小目标的优化,且模型复杂度过高,难以在资源受限的机载平台中应用.为此,本文提出了一种面向无人机载平台的轻量级小目标检测算法YOLOH(You Only Look One Head).首先,针对小目标对基准网络优化,移除深层特征以减少模型参数量,增加浅层特征以获取小目标信息.其次,在特征融合部分加入NAM注意力,增强对小目标的感知能力.接着,设计了多感受野聚焦模块MRFF,以挖掘特征图的感受野信息,增强模型的多尺度检测能力.最后,使用LAMP算法对模型剪枝,去除冗余神经元以压缩模型.实验结果表明,与YOLOv8s相比,YOLOH的模型参数量和计算量分别减少了92%和35%,FPS提高了57%.在VisDrone2019和CARPK数据集上AP_(S)分别提高了3.3%和3.7%.与其他轻量级模型相比,所提YOLOH具有最佳的整体性能,同时平衡了模型大小、精度和推理速度,为无人机载平台的目标检测提供了有效的解决方案.Object detection on UAV platforms has significant applications in both military and civilian domains.However,existing detection methods generally focus on multi-scale object detection and lack optimization for small objects,with model complexities that make them challenging to deploy on resource-constrained aerial platforms.To address this,this paper proposes a lightweight small object detection algorithm for UAV platforms called YOLOH(You Only Look One Head).Firstly,the baseline network is optimized specifically for small objects by removing deep features to reduce model parameters and by increasing shallow features to capture information on small objects.Secondly,an NAM attention mechanism is incorporated into the feature fusion module to enhance small object perception.Then,a multi-receptive-field focusing module(MRFF)is designed to extract receptive field information from the feature map,boosting the model's multi-scale detection capability.Finally,the LAMP algorithm is used to prune the model,removing redundant neurons to compress its size.Experimental results show that compared to YOLOv8s,YOLOH achieves a 92%reduction in model parameters,a 35%decrease in computation,and a 57%increase in FPS.On the VisDrone2019 and CARPK datasets,AP_(S)are improved by 3.3%and 3.7%,respectively.Compared with other lightweight models,the proposed YOLOH offers the best overall performance,balancing model size,accuracy,and inference speed,providing an effective solution for target detection on UAV platforms.

关 键 词：机载平台 YOLOH 小目标检测 轻量级 多感受野 

分 类 号：V279[航空宇航科学与技术—飞行器设计] TP391.41[自动化与计算机技术—计算机应用技术]