关键词: 卷积神经网络, 迁移学习, 玉米叶片, 病害识别, 图像分类, 深度学习

Abstract: Using convolutional neural network to recognize and classify crop disease images requires a long model training time. The method of transfer learning can effectively improve the recognition efficiency. This paper firstly explored the identification effect of transfer learning when all network layers were frozen, parts of network layers were frozen and no network layer was frozen. Furthermore, the InceptionV3 model and the Xception model were used to identify and classify healthy maize leaves, maize leaf spot, sheath blight and rust based on the principle of transfer learning. Experimental results showed that the classification efficiency was highest when the transfer learning did not freeze the network layer, with the accuracy being 97.42%. The accuracy of the InceptionV3 model was up to 92.04% when the number of trainable parameters was set to 70%, while the Xception model reached an accuracy of 94.62% when the number of trainable parameters was 80%. However, the accuracy rate only went up to 87.10% when all the network layers were frozen. Moreover, the Xception model was more suitable than the InceptionV3 model to identify maize leaf diseases.

Key words:  Convolutional neural network, transfer learning, maize leaves, disease identification, image classification, deep learning