Abstract:  To achieve automatic identification of tea diseases on resourcelimited edge devices, a deep learning model deployment method based on edge intelligence is proposed. Firstly, the automated model pruning （AMC） algorithm was used to prune the model of MobileNetV2 on the PlantVillage dataset. Then the model AMC-MobileNetV2 generated at a pruning rate of 90% was used to perform migration learning training on the selfbuilt tea disease dataset. Finally, the obtained tea disease recognition model was deployed on the edge devices. The experimental results show that AMC-MobileNetV2 improves the recognition speed of the model on resourcelimited edge devices with a 94.5% reduction in the number of model parameters and 93.4% reduction in storage volume compared with MobileNetV2, and the average accuracy of the recognition of eight tea diseases is as high as 97.42%. The results of this study can be applied to tea garden disease control robots.

Key words: tea leaf diseases, automatic recognition, transfer learning, automatic pruning, edge intelligence

摘要： 为实现在资源有限的边缘设备上自动识别茶叶病害，提出基于边缘智能的深度学习模型部署方法。首先使用自动化模型剪枝（AMC）算法在PlantVillage数据集上对MobileNetV2进行模型剪枝，然后使用剪枝率为90%时生成的模型AMC-MobileNetV2在自建茶叶病害数据集上进行迁移学习训练，最后将获得的茶叶病害识别模型部署在边缘设备上。试验结果表明，AMC-MobileNetV2与MobileNetV2相比，在模型参数量减少94.5%、存储体积减小93.4%的情况下，提高模型在资源有限边缘设备上的识别速度，对8种茶叶病害识别平均准确率高达97.42%。研究结果可应用于茶园病害防治机器人。

关键词: 茶叶病害, 自动识别, 迁移学习, 自动化模型剪枝, 边缘智能