Distilled-mobilenet model of convolutional neural network simplified structure for plant disease recognition.
Distilled-mobilenet model of convolutional neural network simplified structure for plant disease recognition.
复制标题
DOI：
10.12133/j.smartag.2021.3.1.202009-sa004
复制
发表时间：
2021-01-01
期刊：
影响因子：
--
通讯作者：
Yi, W. M.
中科院分区：
文献类型：
--
作者：
Qiu, Wen-jie;Ye, Jin;Yi, W. M.
卷积神经网络（CNN）的发展带来了大量的网络参数和庞大的模型体积，极大地限制了其在计算资源较小的设备上的应用，如单片机和移动的设备。为了解决这一问题，本文研究了一种结构化模型压缩方法。其核心思想是利用知识蒸馏将复杂集成模型中的知识转移到轻量级的小规模神经网络中。首先，使用VGG 16训练一个具有较高识别率的教师模型，其体积远大于学生模型。然后通过蒸馏将模型中的知识转移到移动网络中。VGG 16模型的参数数大大减少。将知识提取模型命名为Distilled-MobileNet，并应用于38种常见病害（白粉病、黄龙病等）的分类任务。14种作物（大豆、黄瓜、番茄等）。在VGG 16、AlexNet、GoogleNet和ResNet四种不同网络结构上进行的知识蒸馏性能测试表明，当使用VGG 16作为教师模型时，模型的准确率提高到97.54%。在真实的环境中，以单病种识别率、平均准确率、模型记忆和平均识别时间4个指标对训练后的Distilled-MobileNet模型进行准确率评价，结果表明，模型的平均准确率达到97.62%，平均识别时间缩短至0.218 s，仅占VGG 16模型的13.20%，模型大小仅为19.83 MB，比VGG 16小93.60%。与传统的神经网络相比，蒸馏移动模型在缩小网络规模、缩短识别时间方面有显著的改进，为在内存和计算资源有限的设备上进行疾病识别提供了新的思路。
The development of convolutional neural networks(CNN) has brought a large number of network parameters and huge model volumes, which greatly limites the application on devices with small computing resources, such as single-chip microcomputers and mobile devices. In order to solve the problem, a structured model compression method was studied in this research. Its core idea was using knowledge distillation to transfer the knowledge from the complex integrated model to a lightweight small-scale neural network. Firstly, VGG16 was used to train a teacher model with a higher recognition rate, whose volume was much larger than the student model. Then the knowledge in the model was transfered to MobileNet by using distillation. The parameters number of the VGG16 model was greatly reduced. The knowledge-distilled model was named Distilled-MobileNet, and was applied to the classification task of 38 common diseases (powdery mildew, Huanglong disease, etc.) of 14 crops (soybean, cucumber, tomato, etc.). The performance test of knowledge distillation on four different network structures of VGG16, AlexNet, GoogleNet, and ResNet showed that when VGG16 was used as a teacher model, the accuracy of the model was improved to 97.54%. Using single disease recognition rate, average accuracy rate, model memory and average recognition time as 4 indicators to evaluate the accuracy of the trained Distilled-MobileNet model in a real environment, the results showed that, the average accuracy of the model reached 97.62%, and the average recognition time was shortened to 0.218 s, only accounts for 13.20% of the VGG16 model, and the model size was reduced to only 19.83 MB, which was 93.60% smaller than VGG16. Compared with traditional neural networks, distilled-mobile model has a significant improvement in reducing size and shorting recognition time, and can provide a new idea for disease recognition on devices with limited memory and computing resources.