| 摘 要: 为了提高马铃薯叶片病害检测效率以及降低模型在资源受限设备上的部署要求,提出一种基于改进YOLOv8n的马铃薯叶片病害检测模型YOLOv8-VSG。具体改进如下:首先,引入轻量级神经网络架构VanillaNet模块替换原模型的骨干特征提取网络,有效缩减了模型体积;其次,在Backbone嵌入轻量化注意力机制SimAM(Simple Attention Module),增强模型对不同病害区域的关注度,提升网络检测性能;最后,使用GIoU(Generalized Intersection over Union)为边界框回归损失函数,加快模型收敛速度,提升模型泛化能力。实验结果表明,YOLOv8-VSG模型在mAP@0.5和mAP@0.5∶0.9指标上分别达到93.4%和91.5%,较YOLOv8n模型分别提升了2.2%和2.1%,同时检测速度提高了32.9%,参数量减少了37.2%,该方法为马铃薯叶片病害快速、准确检测及实现端侧设备部署提供了参考。 |
| 关键词: 目标检测 YOLOv8n 马铃薯叶片病害 轻量化 |
|
中图分类号: TP391.4
文献标识码: A
|
| 基金项目: 甘肃省高等学校创新基金项目(2022B-107,2019A-056) |
|
| Lightweight Potato Leaf Disease Detection Method Based on Improved YOLOv8n Model |
|
CAI Yaling, KANG Lijun, DAI Yongqiang
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
207729568@qq.com; 42223162@qq.com; dyq@gsau.cdu.cn
|
| Abstract: To enhance potato leaf disease detection efficiency and reduce model deployment requirements on resource-constrained devices, an improved YOLOv8n-based model named YOLOv8-VSG is proposed. Specific improvements are as follows: Firstly, the VanillaNet module replaces the original backbone feature extraction network
to reduce model size. Secondly, the SimAM is embedded into the Backbone to strengthen the model's focus on different disease regions and improve detection performance. Finally, GIoU is adopted as the bounding box regression loss function to accelerate convergence and enhance generalization ability. Experimental results show that the YOLOv8-VSG model achieves 93.4% mAP@ 0.5 and 91.5% mAP@ 0.5:0.9, outperforming YOLOv8n by 2.2% and 2.1%,respectively. Detection speed was increased by 32. 9% , while parameters was decreased by 37. 2% . This method provides a reference for rapid, accurate potato leaf disease detection and deployment on edge devices. |
| Keywords: object detection YOLOv8n potato leaf disease lightweight |
