作者:黄英来，李宁，刘镇波，张彦华
Authors:HUANG Yinglai,LI Ning,LIU Zhenbo,ZHANG Yanhua摘要:针对传统人工检测葡萄叶片病害种类准确率不高、效率低的问题，提出了一种改进轻量卷积网络模型的葡萄叶片病害图像分类方法。选择对小规模数据分类效果好的Xception网络作为基础模型，对其进行创新性改进，首先，将原始网络中的ReLU激活函数替换为ELU激活函数；其次，设计全新的全连接层，使用全局最大池化层替换原来的全局平均池化层，并对输出层进行改进；再次，在网络中嵌入通道注意力机制。对数据进行预处理，将数据按4∶1的比例划分训练集和测试集。为模拟现实拍摄情况，在训练时采用数据增强方法，对数据进行扩容，然后将在ImageNet上预训练好的权重参数迁移到改进的模型中。实验结果表明，改进的葡萄叶片病害分类模型（Grape-Xception）的准确率相较于原始模型提高了2.95个百分点达到99.57%，研究模型的规模为81.38MB。与其他网络模型相比，准确率大幅提高，为葡萄叶片病害的快速诊断和及时防控提供了一种准确高效的方法。
Abstract:Aiming at the problem of low accuracy and efficiency of traditional manual detection of types of grape leaf diseases, a grape leaf disease image classification method with improved lightweight convolutional network model is proposed First, the ReLU activation function in the original network is replaced by the ELU activation function; second, a new fully connected layer is designed, and the global average pooling layer is replaced by the global maximum pooling layer, and the output layer is improved; and then the channel attention mechanism is embedded in the network. The data is preprocessed, and divided into training and test sets in the ratio of 4 In order to simulate realistic shooting situations, methods of data enhancement are used in training to expand the data, and then the weight parameters pretrained on ImageNet are migrated to the improved model The experimental results show that the accuracy of the improved grape leaf disease classification model (Grape-Xception) improved by 2.95 percentage points to 99.57% compared to the original model, and the size of the model was 81.38 MB.Compared with other network models, the accuracy of the model is substantially improved, and it provides an accurate and efficient method for rapid diagnosis and timely prevention and control of grape leaf diseases.

PDF全文下载地址:

可免费Download/下载PDF全文