基于YOLOv8n的鄱阳湖轻量级鸟类目标检测与识别模型研究  

The Research on Lightweight Bird Object Detection and Recognition Modelin Poyang Lake Based on YOLOv 8 n

作　　者：王青羽 姚国清[1] 方朝阳[2] WANG Qingyu;YAO Guoqing;FANG Chaoyang(School of Information Engineering,China University of Geosciences Beijing,Beijing 100083,China;Key Laboratory of Poyang Lake Wetland and Watershed Research,Ministry of Education,Jiangxi Normal University,Nanchang Jiangxi 330022,China)

机构地区：[1]中国地质大学(北京)信息工程学院,北京100083 [2]江西师范大学鄱阳湖湿地与流域研究教育部重点实验室,江西,南昌330022

出　　处：《江西师范大学学报(自然科学版)》2025年第1期86-94,共9页Journal of Jiangxi Normal University(Natural Science Edition)

基　　金：国家自然科学基金(62402461);中央高校基本科研业务费课题(2-9-2022-062)资助项目。

摘　　要：当前鄱阳湖地区鸟类野外自动监测设备的资源有限,致使在野外场景下鸟类快速精准识别存在目标特征不明显、轮廓模糊、尺寸较小等挑战.为了解决这类问题,该文提出了一种基于YOLOv8n的轻量级鸟类检测识别模型YOLOv8-Birds.首先,重新构建模型网络结构,删除深层下采样模块,增加小目标层,以减小模型体量和提升浅层特征权重;其次,融入第3代可变形卷积(DCNv3)设计了C2f_D3模块,提高模糊目标的识别精度;再次,引入分组混洗卷积(GSConv)和加权融合拼接(Concat_BiFPN)模块对颈部网络优化,增强模型特征表达能力,适应不同尺寸目标;最后,应用Slide Loss函数强化困难样本学习.该文以鄱阳湖地区10种珍稀鸟类为研究对象开展模型试验,实验结果表明:精度均值mAP@0.50、mAP@0.75、mAP@0.50∶0.95分别达到93.7%、84.9%、72.8%,测试集鸟类目标平均的正检率提升2.3%,达到89.0%,模型的参数量、体积仅为原模型的50.0%左右.Considering the limited resources of automatic bird monitoring equipment in the Poyang Lake area,rapid and accurate identification of birds in wild scenes presents challenges such as unclear target features,blurred outlines,and small size.Therefore,the lightweight bird detection and recognition model YOLOv8-Birds based on YOLOv8n is proposed.Firstly,the model network structure is reconstructed,the depth downsampling module is removed and a small target layer is added to reduce the model size and enhance shallow feature weights.Secondly,combined with the third generation deformable convolution(DCNv3),the C2f_D3 module improves the recognition accuracy of fuzzy targets,the group shuffled convolution(GSConv)and weighted fusion cascade(Concat_BiFPN)modules are introduced to optimize the neck network,which enhances the feature expression ability of the model,and adapts to targets of different sizes.Finally,the Slide Loss function is applied to enhance learning of difficult samples.The model experiments are conducted on 10 rare bird species in the Poyang Lake area.The experimental results show that the average accuracy rates mAP@0.50,mAP@0.75 and mAP@0.50∶0.95 reach 93.7%,84.9%and 72.8%respectively.The average positive detection rate of bird targets in the test set increase by 2.3%,reaching 89.0%.The number of parameters and volume of the model are only about 50.0%of the original model.

关 键 词：鄱阳湖 鸟类检测识别 YOLOv8n 网络结构优化 可变形卷积 Slide Loss函数 

分 类 号：Q958.1[生物学—动物学] TP391.41[自动化与计算机技术—计算机应用技术]