摘要
为解决复杂环境下茶叶病害识别存在的准确率低、泛化能力差、实时性不足及背景干扰敏感等问题,提出了一种基于改进YOLOv11n的轻量化茶叶病害检测算法——YOLO-tea。该方法在基线模型YOLOv11-nano的基础上,引入了StarNet改进的C3K2模块,有效减少了参数量和计算量,同时增强了多尺度特征提取能力;结合轻量型卷积GSConv构建轻量化颈部网络,减少了参数堆叠,提高了检测精度和推理速度;采用MDPIOU作为边界框损失函数,加速了损失收敛,并精确定位病害目标;最后,通过结合迁移学习策略,增强了模型的泛化能力。实验结果表明,YOLO-tea算法在保持较高检测精度的同时实现了模型轻量化,特别适用于复杂自然环境下茶叶病害的实时精准识别。
In order to solve the problems of low accuracy,poor generalization ability,lack of real-time and sensitive background interference of tea disease identification in complex environment,YOLO-tea,a lightweight tea disease detection algorithm based on improved YOLOv11n,is proposed.Based on the baseline model YOLOv11-nano,this method introduces the improved C3K2 module of StarNet,which effectively reduces the amount of parameters and calculation,and enhances the ability of multi-scale feature extraction.Combining with the lightweight convolutional GSConv to construct the lightweight neck network,the parameter stacking is reduced,and the detection accuracy and reasoning speed are improved.MDPIOU is used as the boundary box loss function to accelerate the loss convergence and accurately locate the disease target.Finally,by combining the transfer learning strategy,the generalization ability of the model is enhanced.The experimental results show that the YOLO-tea algorithm can maintain high detection accuracy and realize the lightweight of the model,which is especially suitable for the real-time accurate identification of tea diseases in complex natural environment.
作者
廖赵明
LIAO Zhaoming(School of Computer Science,Yangtze University,Jingzhou,Hubei 434023,China)
关键词
茶叶病害检测
轻量化
迁移学习
tea disease detection
lightweight
transfer learning
