计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (23): 178-184.DOI: 10.3778/j.issn.1002-8331.2105-0321

• 模式识别与人工智能 • 上一篇    下一篇

改进残差网络在玉米叶片病害图像的分类研究

黄英来,艾昕   

  1. 东北林业大学 信息与计算机工程学院,哈尔滨 150040
  • 出版日期:2021-12-01 发布日期:2021-12-02

Research on Classification of Corn Leaf Disease Image by Improved Residual Network

HUANG Yinglai, AI Xin   

  1. College of Information and Computer Engineering, Northeast Forestry University, Harbin 150040, China
  • Online:2021-12-01 Published:2021-12-02

摘要:

针对传统的玉米叶片病害图像识别方法正确率不高、速度慢等问题,提出一种基于改进深度残差网络模型的玉米叶片图像识别算法。提出的改进策略有:将传统的ResNet-50模型第一层卷积层中7×7卷积核替换为3个3×3的卷积核;使用LeakyReLU激活函数替代ReLU激活函数;改变残差块中批标准化层、激活函数与卷积层的排列顺序。进行数据预处理,将训练集与测试集的比例划分为4∶1,采用数据增强的方式对训练集进行扩充,将改进的ResNet-50模型经过迁移学习得到在ImageNet上预训练好的权重参数。实验结果表明,改进的网络在玉米叶片病害图像分类中得到了98.3%的正确率,与其他网络模型相比准确率大幅提升,鲁棒性进一步增强,可为玉米叶片病害的识别提供参考。

关键词: 玉米叶片病害, 图像识别, 深度学习, 残差网络, 迁移学习

Abstract:

Aiming at the problems of low accuracy and slow speed of traditional corn leaf disease image recognition methods, a corn leaf image recognition algorithm based on improved deep residual network model is proposed. Here are the proposed improvement strategies:replacing the 7×7 convolution kernel in the first convolution layer of the traditional ResNet-50 model with three 3×3 convolution kernels; using the LeakyReLU activation function to replace the ReLU activation function; changing the order of the batch normalization layer, activation function and convolutional layer in the residual block. Firstly, data preprocessing is carried out, dividing the ratio of training set and test set to 4∶1, using data enhancement to expand the training set. Subsequently, the improved ResNet-50 model is subjected to transfer learning to obtain the weight parameters pre-trained on ImageNet. The experimental results show that the improved network has a 98.3% correct rate in corn leaf disease images classification. Compared with other network models, the accuracy rate is greatly improved, and the robustness is further enhanced, which can provide a reference for the recognition of corn leaf diseases.

Key words: corn leaf disease, image recognition, deep learning, residual network, transfer learning