湖北农业科学 ›› 2026, Vol. 65 ›› Issue (11): 194-205.doi: 10.14088/j.cnki.issn0439-8114.2026.11.033

• 信息工程 • 上一篇    下一篇

基于特征解耦下采样机制改进YOLOv11n的轻量化草莓病害检测方法

景鹏斌1, 陈锲2, 于鑫2, 胡静波2   

  1. 1.陕西机电职业技术学院电子信息学院,陕西 宝鸡 721000;
    2.宝鸡文理学院电子电气工程学院,陕西 宝鸡 721000
  • 收稿日期:2026-08-24 发布日期:2026-11-17
  • 通讯作者: 胡静波(1979-),男,陕西渭南人,教授,博士,主要从事机器视觉及嵌入式系统研究,(电子信箱)joyboble@126.com。
  • 作者简介:景鹏斌(1983-),男,陕西渭南人,副教授,硕士,主要从事电子信息技术、嵌入式系统研究,(电子信箱)254597613@qq.com
  • 基金资助:
    陕西省重点研发计划项目(2024NC-YBXM-198); 农业农村部农业物联网重点实验室开放课题(2023AIOT-03); 陕西省职业技术教育学会 2025 年度职业教育教学改革研究课题(2025SZX788); 宝鸡文理学院第十八批教学改革项目(23JGZD15)

Lightweight strawberry disease detection method based on YOLOv11n improved by feature decoupling downsampling mechanism

JING Peng-bin1, CHEN Qie2, YU Xin2, HU Jing-bo2   

  1. 1. Department of Electronic Information, Shaanxi Institute of Mechatronic Technology, Baoji 721000,Shaanxi, China;
    2. School of Electronic & Electrical Engineering, Baoji University of Arts and Sciences, Baoji 721000,Shaanxi, China
  • Received:2026-08-24 Online:2026-11-17

摘要: 针对设施草莓病害检测过程中人工识别效率低以及目标检测模型计算复杂度较高的问题,提出一种基于特征解耦下采样机制的轻量化检测模型DS-YOLOv11n。针对部分轻量化方法侧重卷积算子替换、而缺少结构冗余分析与优化验证的问题,构建“冗余诊断-关键层定位-特征解耦-性能验证”的轻量化优化流程,对YOLOv11n颈部冗余下采样层进行定向改进。通过逐层计算分析确定颈部第17层和第20层为优化层,并将标准下采样卷积解耦为空间特征提取与通道信息融合两个独立过程,以减少空间-通道耦合计算造成的冗余开销。基于包含2 500张设施草莓图像、5 707个标注实例及7类草莓病害目标的数据集,从理论分析、关键层量化评估、消融试验、模型对比和可视化分析等方面对所提模型进行验证。结果表明,与YOLOv11n基线模型相比,DS-YOLOv11n参数量由2.58×106降低至2.42×106,降幅为6.2%;浮点运算量由6.32×109降低至6.11×109,降幅为3.3%;mAP@0.5达到83.71%,较基线模型提升1.34个百分点;mAP@0.5:0.95为61.84%,较基线模型下降0.15个百分点。消融试验表明,仅对颈部关键下采样层进行解耦优化能够在降低模型复杂度的同时保持较高的检测性能,而对骨干网络下采样结构进行全面替换会造成特征表达能力降低。多模型对比结果进一步表明,DS-YOLOv11n在保持较高检测性能的基础上,实现了参数规模和理论浮点运算量的同步降低,为设施农业场景下轻量化目标检测模型设计提供参考。

关键词: 目标检测, 轻量化网络, YOLOv11n, 特征解耦下采样

Abstract: To address the problems of low efficiency of manual identification and high computational complexity of object detection models in facility strawberry disease detection,a lightweight detection model, DS-YOLOv11n, was proposed based on a feature-decoupled downsampling mechanism. Given that some lightweight methods mainly focused on convolutional operator replacement and lacked systematic structural redundancy analysis and optimization evaluation,a lightweight optimization framework consisting of redundancy diagnosis, key-layer localization, feature decoupling, and performance validation was established to selectively optimize redundant downsampling layers in the neck of YOLOv11n. Through layer-wise computational analysis, the 17th and 20th layers in the neck were identified as optimization layers. The standard downsampling convolutions were decoupled into two independent processes, namely spatial feature extraction and channel information fusion, thereby reducing structural redundancy caused by spatial-channel coupled computation.A dataset containing 2 500 images of facility-grown strawberry, 5 707 annotated instances, and 7 strawberry disease categories was constructed to evaluate the proposed model. Comprehensive experiments were conducted from the perspectives of theoretical analysis, key-layer quantitative evaluation, ablation experiments, comparative analysis, and visualization. The results showed that, compared with the YOLOv11n baseline model, DS-YOLOv11n reduced the number of parameters from 2.58×106 to2.42×106, corresponding to a reduction of 6.2%, and reduced floating-point operations (FLOPs)from 6.32 ×109 to 6.11 ×109, with a reduction of 3.3%. Meanwhile, the proposed model achieved an mAP@0.5 of 83.71%, representing an improvement of 1.34 percentage points over the baseline model, while the mAP@0.5:0.95 remained at 61.84%, with only a slight decrease of0.15 percentage points. Ablation experiments demonstrated that selectively applying decoupled optimization to the key downsampling layers in the neck could reduce model complexity while maintaining detection performance, whereas replacing all downsampling structures in the backbone resulted in degraded feature representation capability. Comparative experiments with other lightweight models further indicated that DS-YOLOv11n achieved simultaneous reductions in parameter size and theoretical computational cost while maintaining relatively high detection performance. The proposed method provided a reference for designing lightweight object detection models in facility agriculture applications.

Key words: object detection, lightweight network, YOLOv11n, feature decoupled downsampling

中图分类号: