Classification of Plant Leaf Diseases Using Few Shot Learning

Abstract

Our research on plant leaf disease classification is summarized in this report, emphasizing the use of few-shot learning methods in the task. Although the efficacy of traditional deep learning techniques is prominent, their dependence on large-scale labeled datasets presents a difficulty where the datasets are in short supply. Few-shot learning is shown to be an effective solution in scenarios where there is a lack of labeled data. In the project’s early phases, several feature extractor architectures, such as MaxViT, ConvNeXt, and Swin Transformer have been tested. Afterwards, we focused on making our model more efficient by using smaller and simpler architectures. We found a combination of MobilenetV2, MobilenetV3_small and MobilenetV3_large performs better with increased efficiency. Our experiments with Dense, LSTM and Bi-LSTM classifiers showed that Bi-LSTM providing good results with the lab images, but low accuracy with the field images. The limitations of Bi-LSTM has mitigated with attention mechanism to some extend. In this study, we contribute to the field of plant leaf disease classification by demonstrating the effectiveness of few-shot learning methods, particularly in data-scarce environments. Our work leverages advanced feature extractors like mobilenet v2, and mobilenet v3, and highlights the superior performance of a Bi-LSTM model integrated with an attention mechanism. Following this approach we got an accuracy of 98.23 ± 0.33% in dataset containing laboratory images and 69.28 ± 1.49% in dataset that contains images taken directly from the field. The approach shows significant accuracy improvements, especially in challenging one-shot scenarios, establishing a robust foundation for practical applications on resource-constrained embedded systems.