| Abstrakt: |
Traditional disease identification methods are time-consuming and necessitate specialized knowledge, making them unsuitable for large-scale crop production. In such circumstances, conventional methods of deep learning are impractical. Machine learning techniques, particularly few-shot learning, can help to overcome these challenges by allowing the development of accurate and efficient plant disease identification models with a small amount of training data. The study proposes a few-shot learning strategy for identifying plant diseases, which involves using a pre-trained model on the ImageNet dataset and refining it with a relatively similar type of dataset PlantCLEF2022 for plant-based features extraction. This fine-tuned model is used to extract embedding from plant leaf images to train and classify few shot scenario of plant disease identification. The classification head of the Convolutional Neural Networks (CNN) model is replaced with a Support Vector Machines (SVM) classifier to train with fewer example images in each class. Experiments are performed on popular PlantVillage dataset having 38 classes and PDD271 dataset having 271 classes. The proposed few-shot learning framework outperformed previous methods for few-shot plant disease classification, achieving an average accuracy of 88.4% at 10-shots and 75.5% at 5-shots on PlantVillage dataset. PDD271 dataset is relatively new in few-shot learning but have larger number of classes and the proposed framework provided an accuracy of 56.3% at single-shot, 67.5% for 3-shots and 74.20% for 5-shots scenario. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink