| 摘 要: 针对番茄成熟度检测中环境复杂导致检测难度大,传统目标检测模型参数多,计算量大的问题,提出基于YOLOv11n的轻量化番茄成熟度检测模型 MSD-YOLOv11n。在主干网络中,设计双层主干网络 MultiBackbone-HD,融合 HGNetV2特征提取能力与深度可分离卷积的轻量化特性,通过多尺度门控注意力机制(MSGA)连接双主干增强特征融合效率;在颈部网络中引入Slim-neck结构,通过 GSConv和 VovGSCSP模块替换标准卷积;在头部网络中,加入高效检测头 Detect-Efficient,集成轻量级卷积与分布焦点损失。实验结果表明,在分级良好的番茄成熟度数据集上,MSD-YOLOv11n模型对比原模型在参数量(Parameters)和浮点运算(GFLOPs)上分别减少48.96%、65.63%,且精确率(Precision)、召回率(Recall)、平均精度(mAP@50)、FPS分别达到了89.5%、85.4%、91.7%、286.78frame/s,优于Faster-RCNN、YOLOv10n、YOLOv8n等常见目标检测模型。MSD-YOLOv11n模型既保持高检测精度,又实现参数量和计算量显著降低。 |
| 关键词: YOLOv11n 轻量化 双主干网络 成熟度检测 番茄 |
|
中图分类号: TP391
文献标识码: A
|
|
| MSD-YOLO:Lightweight Tomato Maturity Detection Model Based on Improved YOLOv11n |
|
LI Qilin, WU Lili
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
lql01218@163.com; wull@gsau.edu.cn
|
| Abstract: In response to the challenges posed by the complex environment in tomato maturity detection, as well as the large number of parameters and computational complexity in traditional object detection models. this study proposes a lightweight tomato maturity detection model MSD-YOLOv11n based on YOLOv11n. In the backbone network, a dual layer backbone network MultiBackbone HD is designed, which combines the feature extraction capability of HGNetV2 with the lightweight feature of depthwise separable convolution. The Mult-i Scale Gated Attention mechanism (MSGA) is used to connect the dual backbones and enhance the efficiency of feature fusion; In the neck network, the Slim neck structure is introduced, and the standard convolution is replaced by GSConv and
VovGSCSP modules. In the head network, an efficient detection head Detect Efficient is added, integrating lightweight convolution and distributed focus loss. Experimental vesults have shown that on a well graded tomato maturity dataset,the MSD-YOLOv111n model reduces the number of Parameters and Giga Floating-Point operations (GFLOPs) by 48.96% and 65.63% , respectively, compared to the original model. Additionally, it reduces Precision, Recall, and mean Average Precision(mAP@ 50)FPS reached 89.5% , 85.4% , 91.7% , and 286.78 frame/s respectively, outperforming common object detection models such as Faster RCNN, YOLOv10n, and YOLOv8n. The MSD-YOLOv11n model maintains high detection accuracy while significantly reducing parameter and computational complexity. |
| Keywords: YOLOv11n lightweight dual layer backbone network maturity detection tomato |
