小麦病害的准确识别有助于早期发现病害并进行及时合理的综合防治，从而减少作物产量损失，降低农药使用量。基于机器视觉的作物病害识别已经取得显著成就。然而，该方法需要专家根据经验在少量图像上手工提取特征，这类特征的数量和表征能力有限，难以全面描述实际田间条件下病害症状的多变性。近年来，基于深度学习的卷积神经网络方法被用于提取海量数据的特征，在植物病害识别领域得到广泛研究。卷积神经网络可以自动从图像上逐层提取局部和全局特征，特征的数量多且表达能力强，可以全面地描述真实环境中的病害信息。本文通过消费级相机获取大量小麦叶部病害（主要包括白粉病、叶锈病、条锈病）和健康照片，构建基于卷积神经网络的小麦叶部病害识别方法。首先提出基于迁移学习的两阶段训练策略，然后改进ResNet18构建了轻量级模型ghost-ResNet18，最后在智能手机上部署实现病害识别移动应用程序，具体结果如下：

（1）针对田间能够采集到的小麦病害数据集规模小，导致卷积神经网络训练出现过拟合问题，提出基于迁移学习的两阶段训练策略。该策略通过预先在辅助数据集Plantvillage训练得到预训练模型，然后通过迁移学习的微调方法，将模型参数迁移到小麦病害数据集，并分两个阶段进行从部分参数到整体参数的重新训练，并用当前最先进的六种卷积神经网络（VGG16, DenseNet121, Inception v3, ResNet50, EfficientNet b6, Mobilenetv2）进行建模，利用独立数据验证模型，用识别准确率、参数量和单张图片识别时间等统计指标比较六种网络的性能。结果表明，通过两阶段训练可以有效解决数据量不足引起的过拟合问题，并提高病害识别准确率。在六种网络结构中，分类准确率最高的是Inception v3（92.53%），但是模型参数量大（83.2 M），在 GPU上识别一张图片需要18.31 ms；而参数量最少的是Mobilenetv2（13.63 M），其识别速度排第二（7.55 ms），但是识别准确率只有86.87%；识别速度最快的是VGG16，原因可能是其较低的网络复杂度（4.64 ms）。其余中大规模卷积神经网络模型参数量为30.87-98 M，单张图片识别时间为12.53-20.31 ms。

（2）针对中大规模卷积神经网络模型参数量大，应用到移动设备时易出现内存消耗大、识别速度慢等问题，本研究提出了一种改进的ResNet18结构。该架构通过ghost卷积改进网络的卷积方式，以减少模型卷积层的参数量，本研究中命名为ghost-ResNet18。并将ghost-ResNet18分别与最先进的、高效的、小规模卷积神经网络(Mobilenetv2、Mobilenetv3、GhostNet、Shufflenet v2)以及改进前的ResNet18进行比较。结果表明，本文提出的ghost- ResNet18网络模型参数量22.06 M，识别准确率93.74%，在GPU上对单张图像的识别只需要4.26 ms。其参数量与原始的ResNet18（42.83 M）相比，ghost-ResNet18参数量减少48.49%。识别准确率与其余五种两阶段训练的卷积神经网络相比，高出0.61%-6.87%。在识别速度上，本文改进的ghost-ResNet18识别时间最少。

（3）为实现利用移动设备进行病害识别的可能性，本研究通过在智能手机设备上部署小麦病害识别模型，形成了一个基于卷积神经网络的Android手机病害识别移动应用程序。即使在没有互联网连接的条件下也可以进行识别，农户不需要任何病害知识就可以利用Android手机应用程序实现智能诊断，进而实时监测病情并采取合理的防治手段。

Accurate identification of wheat diseases is helpful for early detection of diseases and implement reasonable prevention and control measures, thereby reducing crop yield losses and pesticide usage. Recognition of crop diseases based on machine vision has made remarkable achievements. However, this method requires experts to manually extract features from a small number of images based on experience. The number and characterization capabilities of such features are limited, and it is difficult to describe the variability of disease symptoms under actual field conditions. In recent years, the convolutional neural network based on deep learning has been used to extract the features of massive data and has been widely studied in plant disease recognition. The convolutional neural network can automatically extract local and global features layer by layer from the image. The number of features is large and the expression ability is strong, which can comprehensively describe the disease information in the real environment. In this paper, a large number of wheat leaf diseases (mainly including powdery mildew, leaf rust, and stripe rust) and healthy leaf images are obtained through consumer-grade cameras, and a method for identifying wheat leaf diseases based on convolutional neural networks is constructed. First, a two-stage training strategy based on migration learning is proposed, then ResNet18 is improved to construct a lightweight model named ghost-ResNet18, and finally a mobile application for disease recognition is deployed on a smartphone. The specific results are as follows:

(1) To deal with the over-fitting problem of convolutional neural network caused by the small scale of the collected wheat disease dataset in the field, a two-stage training method based on transfer learning is proposed. This strategy obtains a pre-trained model by pre-training on the auxiliary data set Plantvillage, and then transfers the model parameters to the wheat disease data set through the fine-tuning method of transfer learning, which retraining from partial parameters to overall parameters in two stages. The current six advanced convolutional neural networks (VGG16, DenseNet121, Inception v3, ResNet50, EfficientNet b6, Mobilenetv2) are used for modeling, which using independent data to validate the model and using statistic indicators such as recognition accuracy, parameter amount, and single image recognition time to compare the performance of six networks. The results show that the two-stage training can effectively solve the over-fitting problem caused by insufficient data and improve the accuracy of disease recognition. Among the six network structures, the highest classification accuracy is Inception v3 (92.53%), but the model parameter is large (83.2 M), and it takes 18.31 ms to recognize a picture on the GPU. The network with least parameter is Mobilenetv2 (13.63 M), its recognition speed ranks second (7.55 ms), but the recognition accuracy rate is only 86.87%. VGG16 achieves the fastest recognition speed (4.64 ms), which may be due to its lower network complexity. The remaining medium and large-scale convolutional neural network model parameters are 30.87-98 M, and the single picture recognition time is 12.53-20.31 ms.

(2) In view of the large memory consumption and slow recognition speed due to large amount of parameters based on medium and large-scale convolutional neural network, this research proposes an improved ResNet18 structure. This architecture improves the convolution method of the network through ghost convolution to reduce the parameter number of the convolution layer in the model, which is named ghost-ResNet18 in this study. ghost-ResNet18 is compared with the most advanced, efficient, lightweight convolutional neural network (Mobilenetv2, Mobilenetv3, GhostNet, Shufflenet v2) and original ResNet18. The results show that the ghost-ResNet18 network model proposed in this paper has 22.06 M parameters and a recognition accuracy of 93.74%. It only takes 4.26 ms to recognize a single image on the GPU. Compared with the original ResNet18 (42.83 M), the parameter amount of ghost-ResNet18 is reduced by 48.49%. Compared with the other five two-stage training convolutional neural networks, the recognition accuracy is 0.61%-6.87% higher than other those using two-stage training. In terms of recognition speed, the improved ghost-ResNet18 in this article takes the least recognition time.

(3) In order to realize the possibility of disease recognition by mobile devices, this research has formed a mobile application for disease recognition in Android mobile phone based on convolutional neural network, by deploying a wheat disease recognition model on smart phone devices. The application can identify diseases even without Internet connection. Farmers can use the Android mobile phone application to realize intelligent diagnosis without any knowledge of the disease, monitor the disease in real time and adopt reasonable prevention and treatment methods.