Dual-graph knowledge distillation for few-shot class-incremental microorganism recognition.

Area of Science:

• Environmental Science
• Computer Science
• Machine Learning

Background:

• Microorganism recognition is vital for environmental monitoring.
• Few-shot class-incremental learning (FSCIL) addresses challenges in recognizing evolving microorganism categories with limited data.
• High annotation costs make traditional supervised learning difficult for new classes.

Purpose of the Study:

• To develop a unified FSCIL framework for environmental microorganism recognition.
• To enhance the model's ability to learn new classes incrementally while preserving knowledge of old classes.
• To address the problem of catastrophic forgetting in incremental learning scenarios.

Main Methods:

• A contrastive-inspired fine-grained representation learning strategy for initial class learning.
• A prototype rectification mechanism to stabilize incremental class representations using base class semantic structures.
• A dual-graph knowledge distillation framework to preserve instance-level and class-level relational knowledge, guided by an exponential moving average teacher model.

Main Results:

• The proposed method achieved the highest average accuracy of 78.19% on the EMDS-7 dataset.
• It maintained the best final-session accuracy of 65.36%, outperforming state-of-the-art FSCIL methods.
• The framework demonstrated effective mitigation of catastrophic forgetting and robust adaptation to new microorganism categories.

Conclusions:

• The unified FSCIL framework significantly improves environmental microorganism recognition in data-scarce, evolving scenarios.
• The proposed components effectively balance learning new information with retaining existing knowledge.
• This approach offers a robust solution for real-world incremental recognition tasks in environmental monitoring.