基于改进轻量级卷积神经网络MobileNetV3的番茄叶片病害识别

番茄病害的及时检测可有效提升番茄的质量和产量。为实现番茄病害的实时无损伤检测，本研究提出了一种基于改进MobileNetV3的番茄叶片病害分类识别方法。首先选择轻量级卷积神经网络MobileNetV3，在Image Net数据集上进行预训练，将预训练得到的共享参数迁移到对番茄叶片病害识别的模型上并做微调处理。采用相同的训练方法对VGG16、ResNet50和 Inception-V3 三种深度卷积网络模型也进行迁移学习并进行对比，结果显示MobileNetV3的总体学习效果最好，在Mixup混合增强和focal loss损失函数下对10类番茄病害的平均测试识别准确率达到94.68%。在迁移学习的基础上继续改进 MobileNetV3模型，在卷积层引入空洞卷积和感知机结构，采用GLU（Gated Liner Unit）闸门机制激活函数，训练得到最佳的番茄病害识别模型，平均测试的识别准确率98.25%，模型的数据规模43.57 MB，单张番茄病害图像的检测耗时仅0.27 s。经十折交叉验证（10-Fold Cross-Validation），模型的鲁棒性良好。本研究可为番茄叶片病害的实时检测提供理论基础和技术支持。

Identification of Tomato Leaf Diseases Based on Improved Lightweight Convolutional Neural Networks MobileNetV3

Timely detection and treatment of tomato diseases can effectively improve the quality and yield of tomato. In order to realize the real-time and non-destructive detection of tomato diseases, a tomato leaf disease classification and recognition method based on improved MobileNetV3 was proposed in this study. Firstly, the lightweight convolutional neural network MobileNetV3 was used for transfer learning on the image net data set. The network was initialized according to the weight of the pre training model, so as to realize the transfer and fine adjustment of large-scale shared parameters of the model. The training method of transfer learning could effectively alleviate the problem of model over fitting caused by insufficient data, realized the accurate classification of tomato leaf diseases in a small number of samples, and saved the time cost of network training. Under the same experimental conditions, compared with the three standard deep convolution network models of VGG16, ResNet50 and Inception-V3, the results showed that the overall performance of MobileNetV3 was the best. Next, the impact of the change of loss function and the change of data amplification mode on the identification of tomato leaf diseases were observed by using MobileNetV3 convolution network. For the test of loss value, focal loss and cross entropy function were used for comparison, and for the test of data enhancement, conventional data amplification and mixup hybrid enhancement were used for comparison. After testing, using Mixup enhancement method under focal loss function could improve the recognition accuracy of the model, and the average test recognition accuracy of 10 types of tomato diseases under Mixup hybrid enhancement and focal loss function was 94.68%. On the basis of transfer learning, continue to improve the performance of MobileNetV3 model, the dilated convolution convolution with expansion rate of 2 and 4 was introduced into convolution layer, 1×1 full connection layer after deep convolution of 5×5 was connected to form a perceptron structure in convolution layer, and GLU gating mechanism activation function was used to train the best tomato disease recognition model. The average test recognition accuracy was as high as 98.25%, the data scale of the model was 43.57 MB, and the average detection time of a single tomato disease image was only 0.27s, after ten fold cross validation, the recognition accuracy of the model was 98.25%, and the test results were stable and reliable. The experiment showed that this study could significantly improve the detection efficiency of tomato diseases and reduce the time cost of disease image detection.