具有高泛化杂草检测能力的农用智能除草车辆研究
Research on agricultural smart weeding vehicles with high generalization capability for weed detection
盖倞尧 1刘世杰 1潘荥春 2潘明章1
作者信息
- 1. 广西大学 机械工程学院,南宁 530004
- 2. 广西柳工农业机械股份有限公司,广西 柳州 545007
- 折叠
摘要
针对现有智能除草机械在杂草检测定位能力不足的问题,设计了一种具有高泛化杂草检测定位能力的农用机械式智能除草车辆.该车辆装备图像传感器采集农田杂草图像,运用深度学习模型进行杂草定位,通过伺服电机控制旋转耙齿盘实现在不伤害作物苗的前提下精确除草.为提高杂草检测的泛化能力,引入了基于对抗生成网络的数据增强策略,通过自注意力机制和梯度惩罚机制优化后的生成网络,提升了杂草图像的生成质量和网络训练的稳定性.通过使用生成田间杂草图像合成新图像以扩充数据集,提高了数据集全面性、多样性.研究结果显示,该智能除草车辆不仅除草效率高,伤苗率低,而且其杂草检测系统的泛化能力得到显著提升.
Abstract
To address the limitations in weed detection and localization capabilities of existing smart weeding machines,we develop a new agricultural smart weeding vehicle with enhanced generalization capabilities for weed detection and localization.Equipped with an image sensor for capturing field weed images,it utilizes a deep learning model for precise weed localization and ensures accurate weeding without harming crops via a servo motor-controlled rotating rake disc.Furthermore,a data augmentation strategy based on generative adversarial networks is implemented,enhancing the generation quality of weed images and the stability of network training.The strategy involves synthesizing new images from field weed images to expand the dataset,improving its comprehensiveness and diversity.Our results show our newly developed smart weeding vehicle not only operates with high efficiency and low crop damage but also boosts the generalization ability of the weed detection system.
关键词
智能车辆/自注意力机制/生成对抗网络/杂草检测Key words
intelligent vehicles/self-attention mechanism/GAN/weed detection引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据