Dual adapter and visual prompt tuning for few-shot leaf disease classification

Pre-trained models have demonstrated considerable potential in few-shot leaf disease classification. However, their direct implementation in few-shot learning scenarios poses two critical challenges: suboptimal adaptability and high computational overhead. To tackle these issues, we present a visual prompt-enhanced dual-adapter few-shot leaf disease classification model termed DVPT (Dual Adapter and Visual Prompt Tuning). Specifically, the model employs visual prompts to enable contextual knowledge transfer to downstream tasks, thereby minimizing backbone network parameters required for task storage and deployment. Additionally, a non-trainable non-parametric adapter designed as a query-key caching model for few-shot datasets is incorporated into the architecture. This adapter forms a dual-adapter mechanism by fusing with original prediction information via residual connections, significantly enhancing feature transfer from pretrained models to downstream tasks. Experimental results indicate that under optimized settings, the model achieves 96.64% accuracy on a 16-shot multi-plant disease classification task. Compared with the full-tuning baseline, the parameter count is drastically reduced (1.18× vs. 24.02×), while outperforming state-of-the-art methods in the field. These results underscore DVPT as an efficient and scalable solution for future plant disease recognition systems.