A New Deep Learning-based Dynamic Paradigm Towards Open-World Plant Disease Detection

Area of Science:

• Computer Vision
• Plant Pathology
• Machine Learning

Background:

• Deep learning enhances plant disease detection but is limited by closed-set, static learning.
• Existing models struggle with unseen disease categories and suffer knowledge degradation when learning new data.

Purpose of the Study:

• Introduce a novel open-world setting for plant disease detection.
• Enable models to infer and dynamically learn previously unseen plant diseases.
• Address knowledge degradation in static learning models.

Main Methods:

• Utilize an unknown-aware region proposal network for pseudo-labeling unknown diseases.
• Employ a class-agnostic classifier to improve recall for unknown diseases.
• Implement a sample replay strategy to preserve knowledge of previously learned classes.

Main Results:

• The proposed method effectively detects both known and unknown plant diseases.
• Demonstrated robust generalization capabilities, including cross-species disease detection.
• Experimental evaluations and ablation studies confirm the method's efficacy.

Conclusions:

• The open-world, dynamically updated approach shows potential as a future paradigm for plant disease detection.
• Identified open issues in classification and localization, proposing future research directions.
• Encourages community research in open-world plant disease detection challenges.