【摘要】 葡萄品种繁多、性状各异，识别葡萄品种速度慢、精度低、成本高，且主观性强、时效性差。因而，开发识别速度快、精度高、成本低、时效性强的葡萄品种识别技术具有重要理论意义和实践价值。为实现葡萄品种的无损、高效识别，为精准农业提供理论基础，以早黑宝、无核早红、夏黑、红地球和阳光玫瑰等5个鲜食葡萄品种为试材，基于其叶片形态特征，采用迁移学习网络模型MobileNet-large，分析该模型在5个葡萄品种上迁移学习的效果，比较3种MobileNet-large网络模型的训练结果，从而构建基于叶片图像的MobileNetV3-large葡萄品种识别模型。结果表明，训练前迁移学习能够显著提高葡萄品种的识别率，无核早红正确识别率可达100%;MobileNetV3-large训练结果的准确率、召回率、F1-score、AUC等因葡萄品种、学习率不同而不同，当学习率为0.005时，MobileNetV3-large模型网络训练损失值最小，其中，红地球葡萄准确率最高。比较3种MobileNet-large网络模型可知，MobileNetV3-large模型整体表现最佳，在训练中第27轮开始收敛，Top-1准确率高达90.56%，平均准确率为97.50%。说明MobileNetV3-large模型是适宜的葡萄品种识别网络模型。更多还原

【Abstract】 Grape(Vitis vinifera L.) has wide varieties and different characteristics, and traditional cultivar identification methods have disadvantages such as slow recognition speed, low accuracy, strong subjectivity, high recognition cost, strong subjectivity, and poor timeliness. Therefore, development of the grape variety recognition technologies with fast recognition speed, high accuracy, low cost, and strong timeliness have important theoretical significance and practical value. To provide a theoretical basis for precision agriculture to realize the non-destructive and efficient identification of grape varieties, in this study,based on the leaf morphological characteristics, five fresh grape varieties including Zaohaibao, Wuhezaohong, Xiahei,Hongdiqiu, and Yangguangmeigui were used as materials, and the transfer learning network model MobileNet-large were used to analyze the effects of transfer learning of the model on five grape varieties, training results of three kinds of MobileNet-large network models were compared, and a MobileNetV3-large grape variety recognition model based on leaf images was constructed. The results showed that transfer learning before training could significantly improve the recognition rate of grape varieties, and the correct recognition rate of Wuhezaohong could reach 100%. The accuracy, recall, F1 score, and AUC of MobileNetV3-large network were different due to different grape varieties and learning rates. The network training loss of MobileNetV3-large model had the smallest at the learning rate of 0.005, and the accuracy of Hongdiqiu was the highest.Comparing three kinds of MobileNet-large network models, the MobileNetV3-large model performed the best overall,converging from the 27th round of training, with a Top-1 accuracy of 90.56%, an average accuracy of 97.50%, indicating that the MobileNetV3-large model was an appropriate identification network model for grape varieties.更多还原