An approach based on performer-attention-guided few-shot learning model for plant disease classification

The evolution of Few-Shot Learning (FSL) technologies has significantly enhanced the capacity for accurate plant disease classification. This paper introduces an FSL model that integrates the Performer-attention mechanism, marking a novel exploration in the domain of plant disease detection. The proposed approach is based on the Performer-attention mechanism that significantly enhances the model's learning efficiency from limited examples and improves disease classification accuracy. Our model is developed through a two-step process: data preprocessing followed by the application of an attention-guided FSL process. This latter step encompasses patch extraction, performer attention, patch embedding, informative patch selection, masked image modeling, and the FSL application. The proposed techniques ensure the capability to address the issue of sample scarcity while ensuring scalability and efficiency. The efficacy of our approach is validated using the PlantVillage dataset and compared with seven state-of-the-art works. Results demonstrate exceptional accuracy rates of 92.15%, 98.12%, and 99.12% across 1-shot, 5-shot, and 10-shot learning scenarios, respectively. These findings depict the potential of our proposed model for more effective crop health monitoring and promoting sustainable agriculture.