基于改进YOLOv8n算法的浆果园内果蝇识别研究  

Research on Fruit Fly Identification in Berry Orchards Based on Improved YOLOv8n Algorithm

作　　者：王威 杨健晟 张梅[1] 陈哲 张群英[2] 刘聂天和 Wang Wei;Yang Jiansheng;Zhang Mei;Chen Zhe;Zhang Qunying;Liu Nietianhe(Electrical Engineering College,Guizhou University,Guiyang 550025,China;Guizhou Botanical Garden,Guiyang 550004,China;Guiyang Huaxi Power Supply Bureau of Guizhou Power Grid Co.,Ltd.,Guiyang 550000,China)

机构地区：[1]贵州大学电气工程学院,贵州贵阳550025 [2]贵州省植物园,贵州贵阳550004 [3]贵州电网有限责任公司贵阳花溪供电局,贵州贵阳550000

出　　处：《山东农业科学》2025年第2期172-180,共9页Shandong Agricultural Sciences

基　　金：国家自然科学基金青年科学基金项目(62003106);贵州省科技支撑计划项目(黔科合支撑〔2022〕一般133,黔科合支撑〔2022〕一般017)。

摘　　要：为了提高浆果园内果蝇的识别效率,以有效指导果蝇防治,本研究以YOLOv8n模型为基准框架,通过结构改进构建了轻量级的果蝇识别算法。具体而言,使用GhostNetV2 bottleneck替代YOLOv8n主干部分所有C2f模块的残差块,构建了全新的C2fGhostV2模块,以降低计算代价并提升识别性能;通过添加卷积层和增加跳跃连接对BiFPN重构,设计了更高效的L-BiFPN结构,替代YOLOv8n颈部的FPN+PAN结构,以提高特征融合效率和表达能力;采用MBConv替代YOLOv8n颈部所有C2f模块的残差块,构建了全新的C2fMBC模块,以提高计算效率并增强对特征的复用能力。实验结果表明,本研究提出的改进YOLOv8n算法的参数、权重和浮点运算次数(FLOPs)比原始YOLOv8n降低48.50%、43.98%和32.10%,精准率、召回率以及平均精确率均值(mAP)分别为97.40%、96.60%和98.32%,明显优于原算法。总体来说,本研究的改进YOLOv8n在显著降低算法复杂度的同时提高了识别精度,具有轻量化和易部署的特性,可以满足浆果园内移动端果蝇识别任务的需求,从而为果农精准防治果蝇提供参考。To improve the identification efficiency of fruit flies in berry orchards to effectively guide fruit fly control,this study proposed a lightweight detection algorithm through structure modification to the YOLOv8n framework.Firstly,a new C2fGhostV2 module was constructed using the GhostNetV2 bottleneck to replace the residual blocks of all C2f modules in the backbone of YOLOv8n,in order to reduce the computational cost and improve the detection performance.Secondly,the BiFPN was reconstructed by adding convolutional layers and increasing jump connections,and a more efficient L-BiFPN structure was designed to replace the FPN+PAN structure in the neck of YOLOv8n,in order to improve the feature fusion efficiency and expression capability.Thirdly,the MBConv was used to replace the residual blocks of all C2f modules in the neck of YOLOv8n,and a new C2fMBC module was constructed,in order to improve the computational efficiency and enhance the ability to reuse the features.The experimental results showed that the parameter amount,weight,and number of floating-point operations(FLOPs)of the improved YOLOv8n were reduced by 48.50%,43.98%and 32.10%compared with the original YOLOv8n algorithm,and its precision,recall and mean average precision(mAP)were 97.40%,96.60%and 98.32%,respectively,which also outperformed the original algorithm.Overall,the improved algorithm put forward in this study significantly reduced the algorithm complexity but enhanced the recognition accuracy,and featured lightweight and easy deployment,so it could meet the requirement of fruit fly recognition task on mobile devices in the berry orchards,thus providing a reference for fruit farmers to precisely control fruit flies.

关 键 词：果蝇识别 YOLOv8n GhostNetV2 BiFPN C2fMBC 

分 类 号：S126[农业科学—农业基础科学] S436.63