Few-shot crop pests and diseases recognition based on adversarial augmentation and task interpolation

Deep learning has made remarkable advances in the identification of crop pests and diseases. However, it requires a significant quantity of labeled data, which increases the burden of dataset collection and annotation. Few-shot learning employs existing information to learn from limited labeled data and rapidly generalize to unknown novel tasks, overcoming the drawback of deep learning, which requires many labeled samples. Therefore, this paper proposes a few-shot crop pests and diseases recognition based on adversarial augmentation and task interpolation. First, the embedding model is used to generate pseudo-labels for few-shot dataset, from which auxiliary tasks are extracted. Subsequently, a few labeled pests and diseases samples in the base task are pseudo-labeled to get the misclassified adversarial samples, and the adversarial tasks are sampled from these adversarial samples to expand the training tasks. Then, the task interpolation is performed on auxiliary tasks and adversarial tasks to mitigate the performance degradation caused by the adversarial task training. Finally, we combine the mixed loss of the base task and the interpolation task to train the model. Extensive experiments on the Plant and Pest dataset and the Plant Village dataset confirm the superior recognition performance of our method in recognizing few-shot pests and diseases.