PlantCaFo: An efficient few-shot plant disease recognition method based on foundation models

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Traditional plant disease recognition methods are costly and sample-scarce.
• Few-shot learning and transfer learning offer solutions but often need domain-specific pretraining.
• Foundation models show promise in few-shot learning scenarios.

Purpose of the Study:

• To explore the potential of foundation models for efficient few-shot plant disease recognition.
• To propose an end-to-end model, PlantCaFo, integrating knowledge from multiple pretrained models.
• To address challenges in data collection and sample scarcity in plant disease identification.

Main Methods:

• Developed PlantCaFo, an efficient few-shot plant disease recognition model based on foundation models.
• Designed a lightweight dilated contextual adapter (DCon-Adapter) for learning new knowledge.
• Utilized a weight decomposition matrix (WDM) for updating text weights within an end-to-end network.

Main Results:

• Achieved 93.53% accuracy on the PlantVillage dataset in a 38-way 16-shot setting.
• Improved accuracy by 6.80% over the baseline on the Cassava dataset (natural environment images).
• Demonstrated significant accuracy increases on an out-of-distribution dataset, validating generalization.

Conclusions:

• PlantCaFo offers a superior approach to few-shot plant disease identification compared to existing models.
• Foundation models can be effectively adapted for agricultural applications like plant disease recognition.
• The proposed DCon-Adapter and WDM contribute to efficient knowledge integration and adaptation.