一种基于改进YOLOv8n-seg的轻量化茶树嫩芽的茶梗识别模型  

A lightweight model for identifying the stalks of tea buds based on the improved YOLOv8n-seg

作　　者：施武 袁伟皓 杨梦道 许高建[1] SHI Wu;YUAN Weihao;YANG Mengdao;XU Gaojian(School of Information and Artificial Intelligence,Anhui Agricultural University,Hefei 230036,China)

机构地区：[1]安徽农业大学信息与人工智能学院

出　　处：《江苏农业学报》2025年第1期75-86,共12页Jiangsu Journal of Agricultural Sciences

基　　金：安徽省高校自然科学研究重点项目(KJ2020A0106);安徽省重大科技专项(202103b06020013);安徽省大学生创新创业计划项目(S202310364126)。

摘　　要：茶树嫩芽茶梗识别对实现茶叶采摘的自动化和智能化具有重要意义。然而,现有的目标检测算法检测茶树嫩芽茶梗存在精度较低、计算量大、模型体积庞大等问题,限制了其在终端设备上的部署。因此,本研究基于YOLOv8n-seg模型,提出一种轻量化的茶树嫩芽茶梗识别模型YOLOv8n-seg-VLS,并在以下3个方面进行了改进:引入VanillaNet轻量化模块替代原有卷积层,以降低模型的复杂程度;在颈部引入大型可分离核注意力模块(LSKA),以降低存储量和计算资源消耗;将YOLOv8的损失函数从中心点与边界框的重叠联合(CIoU)替换为边界框自身形状与自身尺度之间的损失(Shape-IoU),从而提高边界框的定位精度。在采集的茶叶数据集上进行测试,结果表明,改进后获得的YOLOv8n-seg-VLS模型的平均精度值(mAP)方面表现较好,交并比阈值为0.50的平均精度值(mAP_(0.50))为94.02%,交并比阈值为0.50至0.95的平均精度值(mAP_(0.50∶0.95))为62.34%;模型的准确度(P)为90.08%,召回率(R)为89.96%;改进模型的每秒传输帧数(FPS)为245.20帧,模型的大小为3.92 MB,仅为YOLOv8n-seg大小的57.39%。研究结果为后续茶叶智能化采摘装备的研发提供了技术支持。Identifying the stalks of tea buds is of great significance for achieving automated and intelligent tea picking.However,existing object detection algorithms face significant challenges in terms of low detection accuracy,high computational demands,and large model sizes,which collectively limit their deployment on edge devices.To address these challenges,we proposed a lightweight tea stalk detection model,YOLOv8n-seg-VLS,which was based on the YOLOv8n-seg framework.The model incorporated three significant enhancements.First,the VanillaNet lightweight module was introduced to replace traditional convolutional layers,thereby reducing the model's complexity.Second,a large separable kernel attention(LSKA) module was incorporated into the neck section of the network to minimize memory usage and resource consumption.Third,the loss function of YOLOv8 was modified from center intersection over union(CIoU) to shape-and scale-aware intersection over union(Shape-IoU),thereby enhancing the precision of bounding box localization.The experimental results on a collected tea dataset demonstrated that YOLOv8n-seg-VLS achieved a mean average precision(mAP) of 94.02% at mAP_(0.50) and 62.34% at mAP_(0.50∶0.95),with a precision of 90.08% and a recall of 89.96%.In comparison to the original YOLOv8n-seg,the proposed model demonstrated an improvement in frame rate,reaching 245.20 frames per second(FPS).Moreover,the model size was 3.92 MB,which was only 57.39% of the size of YOLOv8n-seg.These results provide technical support for further development of intelligent tea harvesting equipment.

关 键 词：图像识别 茶叶采摘 轻量化模型 YOLOv8n-seg VanillaNet 

分 类 号：TP212[自动化与计算机技术—检测技术与自动化装置] S571.1[自动化与计算机技术—控制科学与工程]