为实现多种类水稻病害的自动识别,采用卷积神经网络对水稻干尖线虫病、白叶枯病、细菌性条斑病等8种水稻叶部病害图像进行识别。将病害图像通过随机旋转以及亮度和对比度随机改变等方法进行样本扩充后,随机划分80%的图像作为卷积神经网络的训练样本,20%的图像作为测试数据。将训练样本直接输入AlexNet网络与LeNet5网络中进行训练,得到AlexNet_model和LeNet_model。在AlexNet网络上采用模糊C均值聚类(FCM)图像处理和在每层激活函数后添加批归一化层(BN)的两种方法对图像进行识别,得到模型FCM_model和BN_model。结合4种模型识别结果及性能评价指标的分析,可知 BN_model的识别效果最佳。BN_model模型的最终测试识别准确率达99.11%,比AlexNet_model、FCM_model和LeNet_model分别提高了0.23个百分点、0.59个百分点和4.43个百分点。该模型识别能力与泛化能力强,为基于卷积神经网络的水稻病害研究提供了参考。

The convolutional neural network is used to identify eight diseases of rice leaf, including dry tip nematode, bacterial leaf blight, bacterial stripe disease and so on, to realize the automatic recognition of multiple rice diseases by computer vision. After the disease images were expanded through random rotation, random change of brightness and contrast, and so on, 80% of the images were randomly divided into training samples and 20% were divided into test data. The training samples were directly input into the AlexNet and LeNet5 networks for training, and the AlexNet and LeNet_models were obtained. FCM_model and BN_model are obtained using two methods of image recognition on AlexNet network: fuzzy C-means clustering image processing and batch normalization layer after activation function of each layer. From the identification results of the four models and the analysis of model performance evaluation indexes, it can be seen that the BN_model has the best recognition effect. The BN_model has a final recognition rate of 99.11%, which is increased by percentage points of 0.23, 0.59,4.43 than AlexNet_model, FCM_model, and LeNet_model, respectively. The model has strong recognition and generalization ability, which provides reference for the research of rice diseases based on convolutional neural network.