Abstract: Associated weeds not only compete with crops for nutrients and water, but also are the intermediate hosts of a variety of diseases and pests, which has become a difficult problem perplexing the efficient production of crops. With the development of deep learning technology, the automatic detection and classification recognition of weeds have been importantly applied in the process of weed removal. Firstly, this paper expounded the hardware requirements and software implementation process of deep learning applied in the process of weed recognition, analyzed the advantages and disadvantages of different hardware used for deep learning, and expounded the establishment, training basic steps such as model evaluation and model deployment. The research progress of deep learning method in weed and crop recognition and weed classification recognition was discussed. Then it was pointed out that there was a large demand for deep learning data and there was no universal data set at present. And low recognition accuracy was caused by weeds and crops blocking each other, complex lighting environment, and poor machine operation conditions. Finally, it was pointed out that the research on image and spectral data fusion, modularization of weed recognition model, weed growth prediction and embedded model deployment would become the future research direction of weed recognition method based on deep learning.

Key words: weed recognition, deep learning, convolutional neural network, target detection, semantic segmentation

摘要： 伴生杂草不仅与作物争夺养分和水分，而且还是多种病虫害的中间寄主，成为困扰作物高效生产的难题。随着深度学习技术的发展，杂草的自动检测和分类识别在清除杂草过程中得到重要应用。首先阐述应用于杂草识别过程中深度学习的硬件需求以及软件实现过程，分析用于深度学习不同硬件的优缺点，阐述深度学习模型建立、训练、模型评估以及模型部署等基本步骤；并重点论述深度学习方法在杂草和作物识别以及杂草分类识别的研究进展。然后指出深度学习数据需求量大，目前无通用数据集，杂草、作物相互遮挡，光照环境复杂，机器作业条件恶劣等情况下识别准确率低的问题。最后提出图像与光谱数据融合、杂草识别模型模块化、杂草长势预测、模型嵌入式部署研究将成为基于深度学习的杂草识别方法未来的研究方向。

关键词: 杂草识别, 深度学习, 卷积神经网络, 目标检测, 语义分割