基于CNN和迁移学习的农作物病害识别方法研究

互联网是一个巨大的资源库,也是一个丰富的知识库。针对农作物小样本引起的过拟合问题,本研究引入了知识迁移和深度学习的方法,采用互联网公开的ImageNet图像大数据集和PlantVillage植物病害公共数据集,以实验室的黄瓜和水稻病害数据集AES-IMAGE为对象开展相关的研究与试验。首先将批归一化算法应用于卷积神经网络CNN中的AlexNet和VGG模型,改善网络的过拟合问题;再利用PlantVillage植物病害数据集得到预训练模型,在改进的网络模型AlexNet和VGG模型上用AES-IMAGE对预训练模型参数调整后进行病害识别。最后,使用瓶颈层特征提取的迁移学习方法,利用ImageNet大数据集训练出的网络参数,将Inception-v3和Mobilenet模型作为特征提取器,进行黄瓜和水稻病害特征提取。本研究结合试验结果探讨了适用于农作物病害识别问题的最佳网络和对应的迁移策略,表明使用VGG网络参数微调的策略可获得的最高准确率为98.33%,使用Mobilenet瓶颈层特征提取的策略可获得96.8%的验证准确率。证明CNN结合迁移学习可以利用充分网络资源来克服大样本难以获取的问题,提高农作物病害识别效率。

Method for identifying crop disease based on CNN and transfer learning

The internet is a huge resource base and a rich knowledge base. Aiming at the problem of small agricultural samples, the utilization technology of network resources was studied in the research, which would provide an idea and method for the research and application of crop disease identification and diagnosis. The knowledge transfer and deep learning methods to carry out research and experiments on public data sets (ImageNet, PlantVillage) and laboratory small sample disease data (AES-IMAGE) were introduced: first the batch normalization algorithm was applied to the AlexNet and VGG of Convolutional Neural Network (CNN) models to improve the over-fitting problem of the network; second the transfer learning strategy using parameter fine-tuning: The PlantVillage large-scale plant disease dataset was used to obtain the pre-trained model. On the improved network (AlexNet, VGG model), the pre-trained model was adjusted by our small sample dataset AES-IMAGE to obtain the disease identification model of cucumber and rice; third the transfer learning strategy was used for the bottleneck feature extraction: using the ImageNet big dataset to obtain the network parameters, CNN model (Inception-v3 and Mobilenet) was used as feature extractor to extract disease features. This method requires only a quick identification of the disease on the CPU and does not require a lot of training time, which can quickly complete the process of disease identification on the CPU. The experimental results show that: first in the transfer learning strategy of parameter fine-tuning: the highest accuracy rate was 98.33%, by using the VGG network parameter fine-tuning strategy; second in the transfer learning strategy of bottleneck feature extraction, using the Mobilenet model for bottleneck layer feature extraction and identification could obtain 96.8% validation accuracy. The results indicate that the combination of CNN and transfer learning is effective for the identification of small sample crop diseases.