Abstract:  A wide variety and random distribution of weed species made weed identification challenging, resulting in low accuracy and poor real‑time performance. In order to address this issue, a method based on deep convolutional neural networks for identifying spinach weeds was proposed. Initially, a deep convolutional neural network model was used to recognize spinach in segmented grid images, which helped in excluding grid images containing spinach. Subsequently, image processing techniques were applied to segment grid images without spinach, and background grid images lacking green pixels were identified, leaving the remaining grid images marked as weed images. Experimental results showed that the DenseNet model, RegNet model, and ShuffleNet model achieved overall spinach recognition accuracy above 0.965 on the test set, demonstrating excellent identification performance. Regarding recognition speed, the ShuffleNet model exhibited the highest computational efficiency, taking only 14.12 ms to recognize a single original image, corresponding to a frame rate of 70.84 fps, meeting the demands of real‑time weed identification applications. By distinguishing spinach and differentiating weeds from soil, the complexity of weed identification was effectively reduced, while the weed identification accuracy was also enhanced.

Key words: vegetable, weed identification, convolutional neural network, deep learning, color segmentation

摘要： 杂草种类繁多且分布随机导致杂草识别难度大、准确率低和实时性差。为此，提出一种基于深度卷积神经网络的青菜杂草识别方法。首先，利用深度卷积神经网络模型对切分后的网格图像进行青菜识别，以此排除青菜网格图像。进而运用图像处理技术对不包含青菜的网格进行图像分割，识别出不含绿色像素的背景网格图像，并将剩下的网格图像标记为杂草图像。试验结果表明：DenseNet模型、RegNet模型和ShuffleNet模型在测试集识别青菜的总体准确率均高于0.965，展现良好的识别效果。识别速度方面，ShuffleNet模型具有最高的计算效率，识别单张原始图像仅耗时14.12 ms，相应帧率为70.84 fps，满足实时杂草识别应用需求。识别青菜进而区分杂草与土壤，有效降低杂草识别的复杂度，同时提高杂草识别准确率。

关键词: 青菜, 杂草识别, 卷积神经网络, 深度学习, 颜色分割