Enhancing anomaly detection in plant disease recognition with knowledge ensemble.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Plant diseases threaten global food security and public health.
• Current plant disease recognition methods are limited to closed-set scenarios, failing to detect novel diseases.
• Real-world applications require systems capable of identifying both known and unknown plant diseases.

Purpose of the Study:

• To extend plant disease recognition to an open-set scenario.
• To benchmark anomaly detection performance of Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and Vision-Language Models (VLMs).
• To propose a novel knowledge-ensemble method for improved open-set plant disease recognition.

Main Methods:

• Benchmarking anomaly detection of CNNs, ViTs, and VLMs under various fine-tuning strategies.
• Developing a knowledge-ensemble method integrating general and domain-specific knowledge.
• Integrating knowledge in logit and feature spaces to enhance model performance.

Main Results:

• The proposed knowledge-ensemble method significantly outperforms existing baselines in open-set plant disease recognition.
• Reduced False Positive Rate at 95% True Positive Rate (FPR@TPR95) from 43.88% to 7.05% for VLMs with 16-shot learning.
• Achieved a further reduction in FPR@TPR95 from 15.38% to 0.71% in the all-shot setting.

Conclusions:

• The knowledge-ensemble approach demonstrates robustness and generalizability across diverse model architectures and training paradigms.
• This method enhances the identification of both known and unknown plant diseases, crucial for practical agricultural applications.
• The study paves the way for more effective and adaptable AI-driven plant disease management systems.