Abstract:

The study aims to improve the automatic recognition of rice pest and disease images and better guide agricultural pest and disease control. Using a combination of transfer learning and ResNet-18 model, we organized open source plant disease data on the internet, and obtained images of 9 rice pests and diseases, including bacterial blight, blast and Tungro, as well as a healthy leaf as the research objects. 11414 cleaned images were selected to establish a dataset for model training, and the 30% dataset was split as the test set. On the basis of six pre trained models such as ResNet-18, GoogLeNet, VGG-16, and MobileNet-v2, a series of parameter adjustments were made to the transfer model. The results show that: (1) under the consistent training parameters, the proposed model ResNet-18 has significantly higher validation accuracy and lowest loss value compared with MobileNet-v2, AxeNet, VGG-16, GoogLeNet, SqueezeNet, and the original ResNet-18 model. The final accuracy of the model is 96.97%. (2) Compared with the original model, the training accuracy of all transferred learning models has been improved significantly, with the improved accuracy ranging from 5.03% to 13.90%. The optimized training model has the characteristics of fast recognition speed and improved accuracy, which can accurately and quickly identify the type of crop disease, providing support for the automatic diagnosis of crop diseases.

Key words: rice, deep learning, diseases and pests, transfer learning, ResNet-18, image recognition