Abstract: In order to solve the problem of low efficiency and low accuracy of apple leaf disease recognition, a study on apple leaf disease recognition model was launched. Four types of images of apple black rot, apple black star, apple rust and apple healthy leaves were obtained by collecting and organizing the network public dataset, 1 200 images were randomly selected to build the datasets, and data enhancement was used to preprocess the data to improve the robustness of the model and reduce the influence of other factors on the model recognition. Combining transfer learning with convolutional neural networks, on the basis of four pretrained models, InceptionV3, Xception, NasNetmobile, and InceptionResNetV2, the fully connected layer is abandoned, and the global average pooling layer is used to complete the feature integration, while the models are finetuned to compare the recognition effects of each model. The training comparison results of the eight models before and after finetuning show that the finetuned InceptionResNetV2 model can achieve 98.83% accuracy on the test set, which can accurately and quickly identify the disease types and provide a reference for related disease identification.

Key words: apple leaf disease, transfer learning, convolutional neural network, image recognition

摘要： 为解决苹果叶部病害识别效率低、精度低的问题，展开苹果叶片病害识别模型的研究。通过收集并整理网络公共数据集，获取苹果黑腐病、苹果黑星病、苹果锈病和苹果健康叶部四类图像，随机选取1200张图像建立数据集，采用数据增强的手段对数据进行预处理，提高模型的鲁棒性，降低其他因素对模型识别的影响。将迁移学习与卷积神经网络相结合，在InceptionV3、Xception、NasNetmobile、InceptionResNetV2这4种预训练模型的基础上，弃用全连接层，通过全局平均池化层来完成特征整合工作，同时对模型采取微调，对比各模型的识别效果。通过微调前后8种模型训练对比结果表明，微调后的InceptionResNetV2模型在测试集上能达到98.83%的准确率，可以准确、快速地识别出病害类型，为相关病害识别提供参考。

关键词: 苹果叶部病害, 迁移学习, 卷积神经网络, 图像识别