| 摘 要: 胡麻生长后期叶茎团簇难以检测,苗期叶茎分明可做研究对象,但现有模型对硬件要求高,难以在低配置设备中实时检测胡麻幼苗目标。为此提出基于改进YOLO11n的轻量级胡麻幼苗目标检测模型,引入幽灵卷积(Ghost Convolution,Ghost Conv),提出C3k2G_Ghost主干网络结构;颈部网络重新设计,构造出幽灵卷积分组(Group Ghost Conv,GrGhConv),提出 GrGh_Ghost颈部网络结构;研究引入高效耦合检测头(Efficient Coupled Detection Head,ECDH)替换原有检测头,用于降参数和稳精度。实验表明:所提模型对比原模型在体积、参数量和GFLOPs上分别降低56.86%、60.56%和65.00%;在FPS上提升60.33%。实现了模型轻量化效果。 |
| 关键词: 幽灵卷积 幽灵分组卷积 高效耦合检测头 轻量化 YOLO11 目标检测 胡麻幼苗 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 教育部人文社会科学研究一般项目(23YJCZH281);上海市哲学社会科学规划课题(2022ZGL010);信息网络安全公安部重点实验室开放课题 |
|
| A Lightweight Model for Flax Seedling Target Detection Based on Improved YOLO11n |
|
XU Yuanhong, HAN Junying
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
2646235041@qq.com; hanjy@gsau.edu.cn
|
| Abstract: During the later growth stages of flax, the clustering of leaves and stems makes detection challenging,whereas the distinct separation of leaves and stems during the seedling stage makes it a suitable research subject.However, existing models have high hardware requirements, making rea-l time detection of flax seedling targets difficult on low-configuration devices. To address this, a lightweight flax seedling target detection model based on improved YOLO11n is proposed. The model incorporates Ghost Convolution (GhostConv) and introduces the C3k2G_Ghost backbone network structure. The neck network is redesigned, constructing Group Ghost Convolution (GrGhConv) and proposing the GrGh_Ghost neck network structure. Finally, an Efficient Coupled Detection Head (ECDH) is introduced.Experiments show that compared to the original model, the proposed model reduces volume, parameter count, and GFLOPs by 56.86% , 60.56% , and 65.00% , respectively, while improving FPS by 60.33% , achieving the goal of model lightweighting. |
| Keywords: ghost convolution group ghost convolution efficient coupled detection head lightweight YOLO11 target detection flax seedlings |
