Detection system for grape leaf diseases based on transfer learning and updated CNN

Abstract: Leaf diseases have a severe threat to the quality and yield of grapes. However, there is often difficulty in the disease detection and low application rate of convolutional neural network (CNN) in a complicated field environment. In this study, an accurate and intelligent detection system was established to realize the strong robustness and real-time performance for grape leaf diseases using transfer learning and an updated CNN model. Firstly, 1990 images were captured for the healthy leaves and five types of infected leaves from field conditions. The combined datasets of PlantVillage and AI Challenger were used to pre-train the VGG16 network for the fully trained parameters. Secondly, the batch normalization, global average pooling layer, and Center Loss function were utilized to modify the structure of the pre-trained VGG16 network, where there was no change in the parameters of the thirteen front convolutional layers and pooling layers. The updated CNN was fine-tuned with the augmented images of grape leaves from field conditions. Stochastic gradient descent (SGD) and adaptive moment estimation (Adam) optimizers were adopted at the initial learning rates of 0.01 and 0.001 in the phase of fine-tuning for experimental comparison. Thirdly, different equilibrium parameters of Center Loss function were utilized in Softmax classification layer for optimal performance. The updated CNN model was also compared with the state-of-the-art models. Finally, the optimal CNN model was deployed in mobile phones to carry out in field condition. The experimental results showed that the updated model using Adam optimizer behaved with a higher accuracy and more stable performance than those using the SGD in the fine-tuning training phase. There were a higher accuracy, a lower loss value, and smaller vibration in the trained model with a small initial learning rate of 0.001 than those with a larger initial learning rate of 0.01, indicating that a smaller learning rate was more reasonable for fine-tuning training. In addition, the accuracy of the model was improved by the equilibrium parameter with a certain range in a Center Loss function. When the equilibrium parameter was set as 0.12, optimal performance of the updated CNN model was achieved at the initial learning rate of 0.001, where the average classification precision was 0.980 0, the recall was 0.980 1, the F1 score was 0.980 1, the average accuracy was 98.02%, and the testing time per image was 0.327 s. The accuracy of updated CNN increased by 2.82%, while the detection time was reduced by 66.8%, and the number of parameters decreased by 83.4%, compared with the original VGG16 network. The comprehensive performance of the updated VGG16 model was also better than that of AlexNet, ResNet50, and Inception v3 models, indicating obvious advantages in the accuracy, weight space occupation, and testing time for the detection of grape leaf diseases. It infers that the Batch normalization layer can speed up the learning process, whereas, the Global average pooling layer without fully connected layers can greatly reduce the number of weight parameters of the model. Center Loss function improved the ability of fine classification. After deployed into smart phone platform, the detection system maintained an accuracy of 95.67% and detection time of 0.357 s per image for the portable and intelligent diagnostics of grape leaf diseases. The transfer learning provided the possibility of quickly acquiring high-performance model under the condition of small datasets. The finding can provide precise guide for the prevention and control of grape diseases in fields.