Classification of Plant Disease from Leaf Images Using Few-Shot Learning

Abstract

To ensure global food security and the overall profit of stakeholders, the importance of correctly detecting and classifying plant diseases is paramount. In this regard, the emergence of Deep Learning-based architectures has provided remarkable performance in plant disease classification in recent times. However, these solutions often require large-scale datasets and high computation resources to learn generalization and achieve state-of-the-art performance. The unavailability of large public datasets in the domain of plant disease classification and low resource constraints of the end-level devices makes the problem even harder. In this regard, we have proposed two solutions to tackle these existing limitations of the Deep Learning-based systems. At first, to ensure the applicability of these solutions in low-end devices we have proposed a lightweight transfer learning-based approach for detecting diseases from tomato leave images. The proposed pipeline utilizes an effective preprocessing method to enhance the leaf images with illumination correction for improved classification. The system extracts features using a combined model consisting of a pretrained MobileNetV2 architecture and a classifier network for effective prediction. Traditional augmentation approaches are replaced by runtime augmentation to avoid data leakage and address the class imbalance issue. Evaluation on tomato leaf images from the PlantVillage dataset shows that the proposed architecture achieves 99.30% accuracy with a model size of 9.60MB and 4.87M floating-point operations, making it a suitable choice for low-end devices. Afterward, to alleviate the dependency on large publicly available datasets, we proposed a pipeline incorporating the concept of Few-shot learning which can predict leaf diseases with only a few samples. The proposed pipeline produces highly general feature embeddings exploiting nine different state-of-the-art CNN-based feature extractors, which are concatenated and passed to a classifier block consisting of a Bi-LSTM layer for effective prediction. We have shown how the concept of `Domain Adaptation' can be utilized in this regard to enhance the representation capability of the feature extractors on unseen classes. The model has been evaluated on the tomato leaf images from the PlantVillage dataset where it achieved promising accuracies of 89.06, 92.46, and 94.07 respectively for 5-shot, 10-shot, and 15-shot classification. Furthermore, an accuracy of 98.09+-0.77 has been achieved 80-shot classification, which is only 1.2% less than state-of-the-art providing 94.5% reduction in the requirement of training data. Experimental findings show that the proposed pipeline has outperformed all the existing works under single-domain, mixed-domain and cross-domain scenarios.