Semi-supervised few-shot learning approach for plant diseases recognition.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Few-shot classification for plant leaf disease recognition is crucial for agricultural yield.
• Existing methods often overlook valuable unlabeled data in agricultural datasets.

Purpose of the Study:

• To develop a semi-supervised few-shot learning approach for accurate plant leaf disease recognition.
• To enhance agricultural protection by effectively utilizing limited labeled samples and abundant unlabeled data.

Main Methods:

• Proposed a semi-supervised few-shot learning framework using the PlantVillage dataset, split into source and target domains.
• Conducted extensive experiments to validate the approach's correctness and generalization across different domain splits and few-shot parameters (N-way, k-shot).
• Employed a confidence interval method for adaptive pseudo-labeling of unlabeled samples within the semi-supervised process.

Main Results:

• The single semi-supervised method achieved an average improvement of 2.8%.
• The iterative semi-supervised method demonstrated an average improvement of 4.6%.
• The proposed methods showed superior performance compared to existing related works.

Conclusions:

• The developed semi-supervised few-shot learning methods effectively improve plant leaf disease recognition accuracy.
• These methods offer a significant advantage by requiring fewer labeled training samples, making them practical for real-world agricultural applications.