LSNet——一种用于无人机图像中小型牲畜目标检测的高效算法  

LSNet——An Efficient Algorithm for Small Livestock Object Detection in UAV Imagery

作　　者：谌文博 谢小伟 王东亮[4] CHEN Wenbo;XIE Xiaowei;WANG Dongliang(School of Surveying and Geoinformation Engineering,East China University of Technology,330013,Nanchang,PRC;Jiangxi Key Laboratory of Watershed Ecological Process and Information(Platform No.2023SSY01051),East China University of Technology,330013,Nanchang,PRC;Key Laboratory of Mine Environmental Monitoring and Improving around Poyang Lake of Ministry of Natural Resources,East China University of Technology,330013,Nanchang,PRC;Key Laboratory of Land Surface Pattern and Simulation,Institute of Sciences and Natural Resources Research,Chinese Academy of Sciences,100101,Beijing,PRC)

机构地区：[1]东华理工大学测绘与空间信息工程学院,南昌330013 [2]东华理工大学江西省流域生态过程与信息重点实验室,南昌330013 [3]东华理工大学自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室,南昌330013 [4]中国科学院科学与资源研究所陆地表层格局与模拟院重点实验室,北京100101

出　　处：《江西科学》2025年第2期253-261,共9页Jiangxi Science

基　　金：国家重点研发计划项目(2021YFD1300501);东华理工大学江西省数字国土重点实验室开放基金项目(DLLJ202206);江西省研究生创新基金项目(YC2023-S560)。

摘　　要：牲畜种群数量调查是卫生防疫、禁牧、休牧、草畜平衡核定等草原管理工作的重要内容,关系到畜牧业现代转型升级和草原的可持续发展。使用无人机进行牲畜种群数量调查,不可避免地出现牲畜小而密集的问题而导致识别错误。针对该问题,提出了一种高效的LSNet算法,该算法增加了一个检测头(P2)来识别浅层特征图中的小物体,并去除了一个检测头(P5)来减少过度下采样的影响。为了捕获高级语义特征,提出了一个大型内核全局注意力空间金字塔池化(LKGSPP)模块,而且集成了双向特征金字塔网络(BiFPN)结构来增强无人机图像中多尺度特征图的融合效果。此外,还利用了中国内蒙古呼伦贝尔陈巴尔虎旗的无人机图像,建立了用于深度学习的放牧牲畜数据集。实验表明,与YOLOv7相比,LSNet的mAP提高了1.87%。LSNet可以成为无人机影像下草原牲畜检测的有利工具。Livestock population surveys are crucial for grassland management tasks such as health and epidemic prevention,grazing prohibition,rotational grazing,and forage-livestock balance assessment.These surveys are directly related to modernization transformation and upgrading of the livestock industry and the sustainable development of grasslands.When using UAVs for livestock population surveys,it’s inevitable to encounter issues such as small and densely packed livestock that often lead to recognition errors.To address this issue,the paper proposes an efficient algorithm called Livestock Network(LSNet).The algorithm incorporates a low-level prediction head(P2)to detect small objects from shallow feature maps,while removing a deep-level prediction head(P5)to mitigate the effects of excessive down-sampling.To capture high-level semantic features,a Large Kernel Global Attention Spatial Pyramid Pooling(LKGSPP)module is proposed.Additionally,a bidirectional feature pyramid network(BiFPN)structure is integrated to enhance the fusion effectiveness of multi-scale feature maps in UAV images.Furthermore,a dataset of grazing livestock for deep learning using UAV images was developed from the Prairie Chenbarhu Banner in Hulunbuir,Inner Mongolia.The experimental results demonstrate that the proposed module significantly improves the detection accuracy for small livestock objects,with the mean Average Precision(mAP)increasing by 1.78%compared to YOLOv7.Thus,the LSNet method is an effective tool for detecting grazing livestock in UAV imagery.

关 键 词：无人机 YOLOv7 目标检测 牲畜种群数量调查 

分 类 号：P23[天文地球—摄影测量与遥感] S818.9[天文地球—测绘科学与技术]