摘要
[目的]针对复杂背景下葡萄叶片病害检测任务中模型准确率与部署效率难以兼顾的问题,提出一种基于改进YOLOv10n的轻量化实时检测模型。通过结构优化与注意力机制增强,在保证检测精度的同时显著降低计算复杂度,以促进模型在移动设备及实际农业场景中的高效部署。[方法]首先,在Backbone网络中设计C2f-HFDRB模块替代原C2f模块,通过将输入特征划分为高、低频两个分支,强化对病害区域高频信息与局部细节特征的建模能力;其次,采用CAA-HSFPN结构替换Neck网络结构,通过精简高计算量组件实现特征金字塔的高效融合;最后,融入TripletAttention注意力模块,通过捕捉跨空间与通道维度的交互依赖关系,精准聚焦于复杂背景下的病害目标区域。[结果]所提模型精确率达到92.0%,提高1.1%;召回率为91.0%;平均精度均值mAP@0.5达到93.3%,提高1.4%。在计算效率方面,模型的计算量较原模型降低60%,降至3.4 GFLOPs;参数量降低0.95 M,仅为1.95 M;实现优异的轻量化特性。[结论]该模型与主流轻量级检测算法相比,在精度-效率权衡方面表现出明显优势,为葡萄叶片病害的实时精准检测及在资源受限环境下的实际应用提供了有效的技术方案与重要参考。
[Objective]To address the challenge of balancing model accuracy and deployment efficiency in grape leaf disease detection under complex backgrounds,this study proposes a lightweight real-time detection model based on an improved YOLOv10n.Through structural optimization and attention mechanism enhancement,the model significantly reduces computational complexity while maintaining detection accuracy,facilitating efficient deployment on mobile devices and in practical agricultural scenarios.[Methods]First,a C2f-HFDRB module was designed in the Backbone network to replace the original C2f module.By splitting input features into high-and low-frequency branches,the modeling capability for high-frequency information and local details of disease regions was enhanced.Second,the CAA-HSFPN structure was adopted to replace the Neck network,achieving efficient feature pyramid fusion by streamlining high-computation components.Finally,the TripletAttention module was integrated to precisely focus on disease target areas in complex backgrounds by capturing cross-dimensional dependencies across spatial and channel dimensions.[Results]The proposed model achieved a precision of 92.0%,an improvement of 1.1%;a recall rate of 91.0%;and a mean average precision(mAP@0.5)of 93.3%,an increase of 1.4%.In terms of computational efficiency,the model's computational load was reduced by 60%to 3.4 GFLOPs,and the number of parameters decreased by 0.95 M to only 1.95 M,demonstrating excellent lightweight characteristics.[Conclusion]Compared with mainstream lightweight detection algorithms,the proposed method exhibits significant advantages in balancing accuracy and efficiency.It provides an effective technical solution and important reference for real-time,accurate detection of grape leaf diseases and practical applications in resource-constrained environments.
作者
乔世成
赵晨雨
李成镛
白明宇
党珊珊
潘春宇
张明月
QIAO Shicheng;ZHAO Chenyu;LI Chengyong;BAI Mingyu;DANG Shanshan;PAN Chunyu;ZHANG Mingyue(College of Computer Science and Technology,Inner Mongolia Minzu University,Tongliao Inner Mongolia 028043,China;Innovation Center for Intelligent Forage Equipment,Inner Mongolia Minzu University,Tongliao Inner Mongolia 028043,China;Shenyang Customas Logistics Management Lenter,Shenyang 110000,China;Tongliao Branch of China United NetworkCommuni-cations Co.,Ltd.,Tongliao Inner Mongolia 028007,China)
出处
《沈阳农业大学学报》
北大核心
2025年第6期45-54,共10页
Journal of Shenyang Agricultural University
基金
国家自然科学基金项目(62162049)
内蒙古民族大学博士科研启动资金项目(BS658)
人兽共患病自治区高等学校重点实验室开放基金项目(MDK2022019)
内蒙古自治区牧草智能装备创新中心开放基金项目(MDK2025050)
内蒙古自治区自然科学基金项目(2025LHMS06012)。
