基于知识蒸馏和轻量级卷积神经网络的植物病害识别方法

Plant disease recognition method based on knowledge distillation and lightweight convolutional neural network

[目的]针对现有植物病害识别模型参数大、田间复杂背景下识别准确率低的问题,提出一种基于改进ShuffleNetV2的轻量化植物病害识别模型LWCR-Net(lightweight crop recognition network)。[方法]在基本特征提取模块中引入残差连接,解决梯度消失问题,使模型能够学习更复杂的特征表示;引入CBAM(convolutional block attention module)注意力机制模块,以增强特征提取和模型利用能力;采用DenseNet121作为教师模型,LWCR-Net作为学生模型,利用知识蒸馏策略对模型进行训练,以进一步提升模型性能。[结果]轻量化植物病害识别模型LWCR-Net大小为2.44 MB;F 1值和准确率分别为95.49%和96.16%,较原模型提高4.89%和3.96%。与DenseNet121、ResNet34、MobileNetV3等其他经典网络对比,LWCR-Net模型不仅达到较高的识别准确率,而且模型参数量较少。[结论]LWCR-Net模型能够实现在田间复杂背景下对植物病害的准确识别,且模型所需内存较小,方便部署到移动端,为植物病害智能诊断提供参考。

[Objectives]Addressing the issue of large model parameters and low recognition accuracy in complex field backgrounds in existing plant disease identification methods,this study proposed a lightweight plant disease recognition model LWCR-Net(lightweight crop recognition network)based on improved ShuffleNetV2.[Methods]The residual connection was introduced into the basic feature extraction module to solve the problem of gradient disappearance,so that the model could learn more complex feature representations.Then,the CBAM(convolutional block attention module)attention mechanism module was introduced to enhance the feature extraction and model utilization ability.Finally,DenseNet121 was used as the teacher model and LWCR-Net was used as the student model,and the knowledge distillation strategy was used to train the model to further improve the performance of the model.[Results]The experimental results showed that the model size of the lightweight plant disease recognition model LWCR-Net was 2.44 MB.The F 1 score and accuracy were 95.49%and 96.16%,respectively,which were 4.89%and 3.96%higher than the original model.Compared with other classical networks such as DenseNet121,ResNet34,MobileNetV3,LWCR-Net model not only achieved high recognition accuracy,but also had fewer model parameters.[Conclusions]The LWCR-Net model proposed in this study can realize the accurate identification of plant diseases under complex field background,and the model requires small memory,making it convenient to deploy on the mobile terminal,and provides reference for the intelligent diagnosis of plant diseases.