Few-shot disease recognition algorithm based on supervised contrastive learning.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Crop diseases significantly reduce yield and quality, impacting agricultural production.
• Accurate and rapid plant disease recognition is crucial for farmers.
• Limited availability of disease samples hinders traditional classifier training.

Purpose of the Study:

• To develop a few-shot plant disease recognition algorithm addressing data scarcity.
• To enhance the accuracy and efficiency of plant disease identification using computer vision.

Main Methods:

• A two-phase approach combining supervised contrastive learning and meta-learning.
• Phase 1: Training a generalized encoder using supervised contrastive learning with ample data.
• Phase 2: Utilizing the encoder for feature extraction and employing meta-learning for few-shot recognition via a nearest-centroid classifier.

Main Results:

• The proposed method surpasses nine other few-shot learning algorithms on the PlantVillage dataset.
• Achieved 79.51% accuracy in few-shot potato leaf disease recognition with only 30 training images.
• Demonstrated that image augmentation strategies significantly impact contrastive learning performance.

Conclusions:

• The algorithm effectively performs few-shot disease recognition, even with small batch sizes, by incorporating label information.
• This approach reduces GPU resource requirements compared to traditional contrastive learning.
• The method offers a promising solution for rapid and accurate plant disease identification in agriculture.