Authors

Abstract

Accurate and real-time weed identification is a key technology for ensuring the efficient operation of field weeding robots. It plays a vital role in reducing pesticide usage, minimising environmental pollution, and protecting the agricultural ecosystem. In response to these demands, this paper proposes a lightweight and high-precision weed recognition method, YOLO-LMSW, tailored for Chinese cabbage field scenarios. Firstly, a lightweight multi-scale convolutional module (LMSConv) is designed, upon which the LMSC3k2 structure is constructed to reduce computational complexity and enhance multi-scale feature extraction capabilities. In terms of network architecture, a lightweight backbone network, LMSNet, is built to significantly reduce model parameters while maintaining detection accuracy. Additionally, a detection head, LMSHead, is designed to further optimise the model structure. To improve localisation accuracy and convergence speed, the Wise-IoU (WIoU) loss function is introduced. Experimental results demonstrate that, compared to YOLOv11n, YOLO-LMSW achieves improvements of 1.2%, 1.0%, and 0.6% in precision, recall, and mAP50 respectively, while reducing the Params and FLOPs by 34.6% and 36.5%, respectively. These results demonstrate its application potential in the actual deployment of field weeding robots.

Abstract in Chinese

精准且实时的杂草识别是保障田间除草机器人高效运行的关键技术，对于减少农药施用、降低环境污染以及保护农业生态环境具有重要意义。针对上述需求，本文提出了一种适用于白菜地场景的轻量化高精度杂草识别方法 YOLO-LMSW。本方法首先设计了一种轻量化多尺度卷积模块（LMSConv），并在此基础上构建了LMSC3k2，用以降低计算复杂度并增强多尺度特征提取能力。在网络架构方面，构建了轻量化主干网络LMSNet，以在保持检测精度的同时显著减少模型参数量。此外，设计了检测头LMSHead，进一步优化模型结构。为提升定位精度与收敛速度，引入了Wise-IoU（WIoU）作为损失函数。试验结果表明，相较于YOLOv11n，YOLO-LMSW在精确率、召回率与mAP50上分别提升了1.2%、1.0%和0.6%，参数量与每秒浮点运算数分别减少了34.6%和36.5%。这些结果展示了其在田间除草机器人实际部署中的应用潜力。