基于YOLOv8S的棉花叶病害识别与多功能机器人系统设计

Design of Multifunctional Robot System for Cotton Leaf Disease Identification Based on Improved YOLOv8S

摘要: 针对棉花叶病害检测中病斑尺度差异大、背景干扰复杂、小目标易漏检及精度与效率难以平衡的问题，文章对YOLOv8s模型进行了改进，构建了YOLOv8s-MSFDS模型。在主干网络中引入MKC-Block模块替换原C2f结构，结合多尺寸卷积核与动态蛇形卷积，提升多尺度病斑特征提取能力；在特征融合层引入频率–空间注意力模块FSA与并行位置感知注意力机制，强化病斑边缘细节特征，增强抗干扰能力；在检测头部分引入多尺度通道注意力模块MTA-Net，优化小目标检测精度并降低冗余计算。实验结果表明，相较于原始YOLOv8s模型，YOLOv8s-MSFDS模型的精确率、召回率、平均精度均值(mAP)分别达91.9%、87.9%、85.1%，参数量仅为5.9 MB，在Jetson Nano平台上推理速度达4.99 FPS，满足实时检测需求。模型在保持轻量化的同时实现了更高的检测精度，有效平衡了性能与效率。搭载于基于ROS构建的履带式机器人平台后，在田间真实场景下表现出良好的泛化性与鲁棒性，可为棉花叶病害智能监测提供硬件与算法一体化的技术支撑。

Abstract: Aiming at the problems of large-scale variation of disease spots, complex background interference, easy missed detection of small targets, and difficulty in balancing accuracy and efficiency in cotton leaf disease detection, the YOLOv8s model was improved in multiple dimensions to construct the YOLOv8s-MSFDS model for cotton leaf disease detection. The MKC Block module was introduced into the backbone network to replace the original C2f structure, and combined with multi-size convolution kernels and dynamic snake convolution to improve the model’s ability to extract features of multi-scale disease spots. The frequency-spatial attention module FSA and parallel position-aware attention mechanism were introduced into the feature fusion layer to strengthen the expression of edge detail features of disease spots and enhance the model’s anti-interference ability in complex field backgrounds. The multi-scale channel attention module MTA-Net was introduced in the detection head to optimize the detection accuracy of small target disease spots and reduce redundant computation, which was adapted to the inference requirements of the hardware end. The experimental results showed that compared with the original YOLOv8s model, the precision, recall, and mean average precision (mAP) of the YOLOv8s-MSFDS model reached 91.9%, 87.9%, and 85.1%, respectively, with only 5.9 MB of parameters, and the inference speed reached 4.99 FPS on the Jetson Nano edge computing platform, meeting the demand of real-time field detection. While maintaining lightweight, the model achieved improved detection accuracy and effectively balanced detection performance and computational efficiency. After being mounted on the tracked robot platform built based on ROS, it exhibited good generalization and robustness in real field scenarios, which can provide integrated hardware and algorithm technical support for the intelligent monitoring of cotton leaf diseases.