卷积神经网络的多尺度改进及其在玉米病害症状识别中的应用

Multi-scale improvement of convolutional neural network and its application in symptom recognition of maize disease

为解决传统卷积神经网络模型训练时间长、参数量大、泛化能力弱等问题,提出了一种基于VGG-16的改进多尺度卷积神经网络模型。用一个叠加卷积层替换VGG-16模型的最后3×3×512卷积层,并进行批归一化处理,提高模型训练速度;用全局池化层替换全连接层,大大减少模型参数总量。利用Plant Village公共数据集(健康玉米叶片、灰斑病、锈病和叶枯病叶片)结合大田试验采集的玉米病害图像数据对改进后模型进行训练和测试,并与常见的传统卷积神经网络模型进行对比。结果表明,模型参数和收敛时间均小于传统卷积神经网络,单一背景下的平均分类识别准确率达99.31%,明显优于传统神经网络模型(VGG-16的90.89%、ResNet-50的93.60%、Inception-V3的94.23%、MobileNet-V2的93.83%和DenseNet-201的95.70%)。同时,利用大田复杂背景病害图片测试新模型的泛化性,识别准确率达98.44%,单张图片测试平均仅需0.25 s。

In order to solve the problems of long training time,large amount of parameters,and weak generalization ability of traditional convolutional neural networks,this paper proposes an improved multi-scale convolutional network model based on VGG-16.The last 3×3×512 convolution layer of the VGG-16 model was replaced with a superimposed convolution layer for batch normalization to increase the model learning rate.In addition,the total amount of model parameters was reduced by replacing the fully connected layer with a global pooling layer.The public Plant Village data set(healthy maize leaves,gray spot disease,rust and leaf blight leaves)combined with field collected maize disease images were used to train and test the improved VGG model,and the results were compared with those of the traditional neural network models.The results show that the convergence time of the improved model is significantly shorter than that of the traditional convolutional neural networks,and the average classification accuracy reached 99.31%,which is much better than those of the traditional neural network models(VGG-1690.89%,Resnet-5093.60%,Inception-V394.23%,Mobilenet 93.83%and DenseNet-20195.70%).Meanwhile,field disease images with complex backgrounds were used to test the proposed method,and the average recognition accuracy could reach 98.44%,taking only 0.25 seconds per image.