A Multi-Task Learning and Transfer Learning-Based Model for Crop Leaf Disease Identification

In recent years, crop leaf diseases have been increasing year by year, and their yield and quality have been impacted as a result, it is crucial to identify crop diseases accurately and rapidly. To address the problems of current crop leaf disease identification methods that require a long time, have low accuracy, a huge number of parameters, and can only perform a single task at a time, a convolutional neural network incorporating multi-task learning and transfer learning is proposed for crop leaf disease identification. The images of wheat and maize leaf diseases are used as research objects and the dataset is augmented by image enhancement techniques. The disease images are transferred for recognition by transfer learning using the pre-trained EfficientNetB7 model on ImageNet. Then the alternate learning method is adopted to retain the parameters of the middle layer, as a shared layer for alternate learning, and retrain the fully connected layer for multi-task learning. Experiments show that the fine-tuned transfer learning model presented in this study performs well with fewer epochs, achieves an average recognition accuracy of 99.43%, and reduces the number of training parameters in the model by about 50% compared to the single-task model. Compared with some traditional networks, it has a better fitting effect and higher average accuracy of the model.