TDR-Net: Tobacco Disease Recognition Using Improved GoogLeNet Network

Rapid identification of tobacco diseases can improve the quality of tobacco production and reduce the resource consumption of tobacco disease prevention and control. Due to the lack of models for the image recognition of tobacco disease, this paper proposes a deep convolutional neural network codenamed TDR-Net, which is based on GoogLeNet. Experiments conducted on an original dataset that is expanded by random flipping, binarization, and other methods. Firstly, to optimize the network activation process, an Elu activation function is introduced. Then, we add local response normalization layers to eliminate the problem of excessive activation value and ensure the robustness of the network. Furthermore, the TDR-Net network also introduces a "Generalization Module", which improves the generalization ability and nonlinear expression ability of the network. Finally, we use Mini-Batch Gradient Descent optimizer to reduce the model error. The results showed that the accuracy of TDR-Net is 4.17% higher than GoogLeNet, reaching 97.44%, which provides a reliable reference for the recognition model of tobacco disease.