Abstract:

Due to the diversity and complex distribution of weeds in bok choy fields， the existing methods for weed identification have the problems of low efficiency， poor accuracy and lack of robustness. This study proposed a method to identify bok choy and weeds based on deep convolutional neural networks， using seedling stage bok choy and their associated weeds as the research objects. Firstly， image processing methods were used to mark images containing green plants， and then a neural network model was used to distinguish bok choy and weeds. In order to investigate the recognition effect of different neural network models， the DenseNet model， GoogLeNet model and ResNet model were used to recognize images containing bok choy or weed images， and the F₁ value， overall accuracy and recognition speed were used as evaluation criteria. The experimental results showed that the 3 neural network models could effectively distinguish bok choy and weeds， and the ResNet model was the optimal model， with an overall accuracy and recognition speed of 97.2% and 78.34 frames·s^-1 on the testing datasets， respectively. The bok choy and weed identification method proposed in this study could effectively reduce the complexity of weed identification， improve the robustness and generalization ability of identification， and laid the foundation for the research on precision weed control in bok choy fields.

Key words: deep learning, convolutional neural network, bok choy recognition, weed recognition

摘要：

针对青菜田间杂草种类繁多且分布复杂导致识别效率低、精度差和稳健性不足等问题，以苗期青菜及其伴生杂草为研究对象，提出了一种基于深度卷积神经网络的青菜和杂草识别方法。首先使用图像处理方法标记出包含绿色植物的图像，进而利用神经网络模型对青菜和杂草进行区分。为探究不同神经网络模型的识别效果，分别选取DenseNet模型、GoogLeNet模型和ResNet模型对图像中包含青菜或者杂草图像进行识别，并以F₁值、总体准确率和识别速度作为评价依据。结果表明，3种神经网络模型均能有效区分青菜和杂草，其中ResNet模型为最优模型，其在测试集的总体准确率和识别速度分别为97.2%和78.34 帧·s^-1。提出的青菜和杂草识别方法可有效降低杂草识别的复杂度，并能够提升识别的稳健性和泛化能力，为青菜田间杂草精准防控的研究奠定基础。

关键词: 深度学习, 卷积神经网络, 青菜识别, 杂草识别