Detection of unknown strawberry diseases based on OpenMatch and two-head network for continual learning

Area of Science:

• Computer Science
• Plant Pathology
• Artificial Intelligence

Background:

• Continual learning in plant disease recognition requires distinguishing known from unknown diseases.
• Open Set Recognition (OSR) and Out-of-Distribution (OoD) detection are deep learning techniques for identifying unknown plant diseases.
• Current OSR and OoD methods often require prior exposure to outlier data during training.

Purpose of the Study:

• To compare two deep learning techniques, a two-head network (OoD) and OpenMatch (OSR), for detecting unknown plant diseases.
• To analyze the impact of modified training procedures on the performance of these methods.
• To evaluate their effectiveness in a fine-grained recognition task using strawberry disease images.

Main Methods:

• The study analyzed a two-head network for OoD detection and the semi-supervised OpenMatch for OSR.
• A comparative experiment was designed by modifying the training procedures of both models for fair comparison.
• An image dataset of eight strawberry diseases was created for experimental evaluation.

Main Results:

• Modified training procedures enabled a direct comparison between the two-head network and OpenMatch.
• Both methods demonstrated reasonable performance after training procedure adjustments.
• The models achieved over 90% accuracy in classifying strawberry diseases and detecting unknown diseases.

Conclusions:

• Accurate detection of unknown plant diseases is crucial for effective continual learning systems.
• Standardized training procedures can facilitate performance comparison between OSR and OoD detection models.
• The findings highlight the potential of these deep learning approaches for enhancing plant disease surveillance.