VCPC: virtual contrastive constraint and prototype calibration for few-shot class-incremental plant disease...

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Deep learning advances plant disease recognition but struggles with real-time, novel disease detection.
• Existing systems are mostly offline and lack adaptability to new disease classes under few-shot conditions.
• Real-time crop surveillance requires incrementally adaptive models for few-shot class-incremental learning (FSCIL).

Purpose of the Study:

• To introduce VCPV (virtual contrastive constraints with prototype vector calibration) for sustainable plant disease classification under FSCIL conditions.
• To develop a method that enables real-time crop surveillance systems to identify novel plant diseases with limited examples.
• To improve the adaptability and accuracy of plant disease recognition models in dynamic environments.

Main Methods:

• The VCPV method involves two phases: base class training and incremental training.
• A virtual contrastive class constraints (VCC) module enhances base class learning and allocates embedding space for new classes.
• A prototype calibration embedding (PCE) module distinguishes new categories from existing ones, optimizing prototype space and recognition accuracy.

Main Results:

• VCPV achieved state-of-the-art accuracy on the PlantVillage dataset in both 5-way 5-shot and 3-way 5-shot settings.
• The method demonstrated promising performance on the CIFAR-100 dataset.
• Visualisation results confirmed the strategy's effectiveness in fine-grained, sustainable disease recognition.

Conclusions:

• VCPV offers a sustainable solution for plant disease classification in few-shot class-incremental learning scenarios.
• The method significantly enhances the ability of AI systems to monitor crops for novel diseases in real-time.
• VCPV holds potential for advancing FSCIL in plant disease monitoring and agricultural applications.