基于改进ResNet18的玉米叶片病害识别方法

Recognition Method of Maize Leaf Disease Based on Improved ResNet18

针对传统玉米病害识别方法准确率差、效率低、难以实现在农田复杂环境下识别等问题,本文提出了一种基于改进的ResNet18的玉米叶片病害识别算法,通过在ResNet18网络的基础上添加通道注意力和空间注意力机制,加强对玉米叶片病害区域特征的提取,提高玉米病害识别的准确率;将残差块中的激活函数替换为SeLU激活函数搭配Alpha Dropout并融入网络中,有效防止网络过拟合,并提高模型的收敛速度;引入随机裁剪分支,增加图像样本数据的多样性,还优化了网络结构,提高了模型的鲁棒性和效率。实验结果表明,所提出的CB-SE-ResNet18网络模型在农田环境下的7种玉米病害数据集中的平均准确率达到94.42%,相比原模型高出了3.71%,平均单幅图识别时间为6.49 ms,模型内存占用仅为32.72 MB。改进后的CB-SE-ResNet18模型具有轻量化、识别速度快、识别精度高等特点,为农田复杂环境下的玉米病害识别提供了参考。

In order to solve the problems of low accuracy,low efficiency,and difficulty in implementing traditional maize disease recognition methods in the complex farmland environments,a novel algorithm based on the improved ResNet18 was proposed in this paper.By incorporating the channel attention and spatial attention mechanisms into the ResNet18 network,the algorithm enhanced the extraction of features related to maize leaf diseases,the accuracy in disease recognition was then improved.Furthermore,the activation function in residual blocks was replaced with the SeLU activation function,combined with the Alpha Dropout,to effectively prevent the overfitting and expedite model convergence.Additionally,a random cropping branch was introduced to augment the diversity of image samples.The network structure was optimized to enhance the model robustness and efficiency.Experimental results demonstrated that,the proposed CB-SE-ResNet18 model achieved an average accuracy of 94.42%in the dataset comprising seven types of maize leaf diseases in real field environments.Moreover,compared to VGG16,ResNet34,Xception,Inception V3,and ResNet18,the proposed model outperformed them by 1.85%,2.53%,0.36%,0.68%,and 3.71%respectively.The average recognition time for a single image was 6.49 ms,and the model memory consumption was only 32.72 MB.The improved CB-SE-ResNet18 model had the characteristics of lightweight,fast recognition speed,high recognition accuracy,and better robustness.This work provided a valuable reference for efficient maize disease recognition in field environments.