Abstract: Aiming at the problems of low efficiency, poor recognition ability, and unpopular application of recogni- tion technology in rice blast that affect food security, an algorithm for accurate recognition of rice blast disease based on the improved VGG16Net model in convolutional neural network is proposed. —VGG16-H. Compared with the traditional VGG network model, VGG16-H reduces the computational load of the computer by reducing the number of convolution kernels and increasing the Dropout layer and Group Normalization layer. The GN layer is basically not affected by the batch size. The RiceLeafs data set of lesion images of rice blast was collected from the Internet, and the RiceLeafs original data set was processed by computer vision and OpenCV to per- form random rotation, random brightness transformation, and random contrast on the data set, so as to expand the number of samples and enhance data. Improve the accuracy of the convolutional neural network, enhance the generalization ability of the network model, and have a certain mitigation effect on the over-fitting phenomenon of the network. The preprocessed data set is randomly divided acc- ording to the ratio of training set:test set=8:2, and used as training set and verification set for model training. Through the training of the Pytorch deep learning platform, the convolutional neural network is used to construct the VGG16-H network model, VGG16Net model, AlexNet model and compared with the support vector machine SVM. The lowest training recognition rate is 96.9% of SVM, and the lowest test recognition rate is SVM 96.3%, while the highest training recognition rate is 99.3% of the VGG16-H model, and the highest test recognition rate is 98.7% of the VGG16-H model. Experimental analysis shows that compared with the traditional network model, the improved VGG16Net model—VGG16-H can achieve a higher recognition rate under the condition of limited computing resources.

Key words: Convolutional Neural Networks, Classification recognition, OpenCV , VGG, Support Vector Machine

摘要： 针对影响粮食安全的水稻稻瘟病中人工识别的效率低、识别能力差、识别技术应用不普及的问题，提出卷积神经网络中一种基于改进式VGG16Net模型对稻瘟病病症精准识别的算法—VGG16-H。相对于传统VGG网络模型，VGG16-H通过减少卷积核数量，增加Dropout层和Group Normalization层，减轻了计算机的计算负荷，GN层基本上不受batch size的影响。从互联网上收集到水稻稻瘟病的病斑图像RiceLeafs数据集，将RiceLeafs原始数据集利用计算机视觉和OpenCV对数据集进行随机旋转、随机亮度变换、随机对比度等处理，使样本数量扩充和数据增强，提高卷积神经网络精准度，增强网络模型的泛化能力，且对网络的过拟合现象有一定的缓解作用。预处理后的数据集按训练集：测试集=8：2的比例随机划分，分别作为训练集、验证集进行模型训练。通过Pytorch深度学习平台训练，使用卷积神经网络构建VGG16-H网络模型、VGG16Net模型、AlexNet模型并与支持向量机SVM进行对比，其中训练识别率最低是SVM的96.9%，测试识别率最低为SVM的96.3%，而训练识别率最高为VGG16-H模型的99.3%，测试识别率最高为VGG16-H模型的98.7%。实验分析表明：改进式的VGG16Net模型——VGG16-H相比于传统网络模型在计算资源有限的情况下也能取得较高的识别率。

关键词: 卷积神经网络, 分类识别, OpenCV, VGG, 支持向量机