一种基于改进卷积神经网络YOLOv8n的黄精种子成熟度识别模型  

A Polygonatum sibiricum Seed Maturity Recognition Model Based on Improved Convolutional Neural Network YOLOv8n

作　　者：冯峰 孔淳 陈诗瑶 马春悦 王琛 郭秀茹 孙博 王志军[1] Feng Feng;Kong Chun;Chen Shiyao;Ma Chunyue;Wang Chen;Guo Xiuru;Sun Bo;Wang Zhijun(College of Information Science and Engineering,Shandong Agricultural University,Taian 271018,China)

机构地区：[1]山东农业大学信息科学与工程学院,山东泰安271018

出　　处：《山东农业科学》2025年第1期156-165,共10页Shandong Agricultural Sciences

基　　金：山东省重大科技创新工程项目“现代果园智慧种植装备与大数据平台研发及示范应用”(2019JZZY010706);山东省自然科学基金面上项目“基于类重叠视角的类不平衡数据分类方法研究”(ZR2023MF098)。

摘　　要：黄精是重要的药食同源品种,在黄精的种植与育种过程中,通常采用基于个人经验的人工方式进行种子成熟度识别,但该方式存在主观性强、精确度不稳定的问题,会在一定程度上影响种植后黄精的产量和质量,进而影响黄精产品的品质和下游产业的经济效益;另一方面,由于黄精种植地块通常较为分散,使得人工识别种子成熟度的方式较为低效。为了解决以上问题,本研究提出一种基于改进卷积神经网络YOLOv8n的黄精种子成熟度识别模型——YOLOv8n-FasterNet-EMA。首先,在模型轻量化方面,通过将YOLOv8n主干网络中原本的卷积替换为FasterNet的PConv卷积结构,与Bottleneck层结合后得到新的c2f-FasterNet模块,从而减小模型的计算量与内存消耗;其次,在模型泛化性能方面,通过使用EMA注意力机制与YOLOv8n颈部网络中检测处的c2f模块结合,提升模型的特征提取能力,进而改善模型的泛化能力。为验证所提模型的性能,在构建的黄精种子成熟度数据集上进行了对比实验,结果表明,与原YOLOv8n模型相比,本研究所提模型在平均识别精度上提升了2.1%,同时模型的参数量降低了21.3%;此外,与SSD、YOLOv5n相比,本研究所提模型在识别精度和速度上也取得一定的提升。因此,YOLOv8n-FasterNet-EMA能有效识别黄精种子的成熟度,这对实现黄精种子成熟度的智能化识别,进而提升黄精育种的质量、改善下游产业的经济效益均具有重要的实际意义。Polygonatum sibiricum is an important medicinal and edible homologous variety.In the process of planting and breeding of P.sibiricum,manual methods based on personal experience are usually used to identify the maturity of P.sibiricum seeds.But this method has problems of strong subjectivity and unstable accuracy,which will affect the yield and quality of subsequent P.sibiricum planting to a certain extent,and thus affect the product quality and economic benefits of the downstream industry of P.sibiricum.On the other hand,due to the scattered planting areas of P.sibiricum,the method of manually identifying maturity was relatively inefficient.To address the above issues,this paper proposed a maturity recognition model for P.sibiricum seeds based on an improved convolutional neural network YOLOv8n that was YOLOv8n-FasterNet-EMA.Firstly,in terms of model lightweighting,a new c2f-FasterNet module was obtained by replacing the original convolution in the YOLOv8n backbone network with PConv convolution structure of FasterNet which was then connected to the Bottleneck layer,thereby reduced the computational complexity and memory usage of the model.Then,in terms of model generalization performance,the feature extraction ability and model generalization ability of the model were improved by combining the EMA attention mechanism with the c2f module at the detection point of the YOLOv8n neck network.To verify the performance of the proposed model,comparative experiments were conducted on the collected P.sibiricum seed maturity datasets.The experiments showed that compared with the original YOLOv8n model,the proposed model improved the average recognition accuracy by 2.1%,while reduced the number of model parameters by 21.3%.In addition,compared with SSD and YOLOv5n,the model proposed in this study achieved certain improvements in both recognition accuracy and speed.Therefore,the proposed model could ef-fectively identify the maturity of P.sibiricum seeds,which is of great practical significance for achieving the intelligent

关 键 词：黄精 种子成熟度识别 卷积神经网络 YOLOv8n FasterNet EMA注意力机制 

分 类 号：S126[农业科学—农业基础科学] S567.23