Zhang Kun, Yuan Bohan, Cui Jingying, Liu Yuyang, Zhao Long, Zhao Hua, Chen Shuangchen
College of Physics and Electronic Engineering, Xinyang Normal University, Xinyang, 464000, China.
College of Agricultural Equipment Engineering, Henan University of Science and Technology, Luoyang, 47100, China.
Sci Rep. 2024 Dec 28;14(1):31168. doi: 10.1038/s41598-024-82529-x.
Tea bud detection technology is of great significance in realizing automated and intelligent plucking of tea buds. This study proposes a lightweight tea bud identification model based on modified Yolov5 to increase the picking accuracy and labor efficiency of intelligent tea bud picking while lowering the deployment pressure of mobile terminals. The following methods are used to make improvements: the backbone network CSPDarknet-53 of YOLOv5 is replaced with the EfficientNetV2 feature extraction network to reduce the number of parameters and floating-point operations of the model; the neck network of YOLOv5, the Ghost module is introduced to construct the ghost convolution and C3ghost module to further reduce the number of parameters and floating-point operations of the model; replacing the upsampling module of the neck network with the CARAFE upsampling module can aggregate the contextual tea bud feature information within a larger sensory field and improve the mean average precision of the model in detecting tea buds. The results show that the improved tea bud detection model has a mean average precision of 85.79%, only 4.14 M parameters, and only 5.02G of floating-point operations. The number of parameters and floating-point operations is reduced by 40.94% and 68.15%, respectively, when compared to the original Yolov5 model, but the mean average precision is raised by 1.67% points. The advantages of this paper's algorithm in tea shot detection can be noticed by comparing it to other YOLO series detection algorithms. The improved YOLOv5 algorithm in this paper can effectively detect tea buds based on lightweight, and provide corresponding theoretical research for intelligent tea-picking robots.
茶芽检测技术对于实现茶芽的自动化和智能化采摘具有重要意义。本研究提出了一种基于改进的Yolov5的轻量级茶芽识别模型，以提高智能茶芽采摘的准确率和劳动效率，同时降低移动终端的部署压力。采用以下方法进行改进：将YOLOv5的骨干网络CSPDarknet-53替换为EfficientNetV2特征提取网络，以减少模型的参数数量和浮点运算量；在YOLOv5的颈部网络中引入Ghost模块，构建Ghost卷积和C3ghost模块，进一步减少模型的参数数量和浮点运算量；将颈部网络的上采样模块替换为CARAFE上采样模块，可在更大的感受野内聚合上下文茶芽特征信息，提高模型检测茶芽的平均精度均值。结果表明，改进后的茶芽检测模型平均精度均值为85.79%，仅有414万个参数，浮点运算量仅为5.02G。与原始Yolov5模型相比，参数数量和浮点运算量分别减少了40.94%和68.15%，但平均精度均值提高了1.67个百分点。通过与其他YOLO系列检测算法进行比较，可以看出本文算法在茶芽检测方面的优势。本文改进的YOLOv5算法能够基于轻量级有效地检测茶芽，为智能采茶机器人提供相应的理论研究。
