Plant and Disease Recognition Based on PMF Pipeline Domain Adaptation Method: Using Bark Images as Meta-Dataset.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Efficient image recognition is crucial for crop and forest management but hindered by numerous species, appearance variations, and limited labeled data.
• Existing methods struggle with the scarcity of training data and the need for domain generalization in plant identification tasks.

Purpose of the Study:

• To enhance a state-of-the-art Cross-Domain Few-shot Learning (CDFSL) method for improved plant and disease recognition.
• To address challenges in image recognition, including data scarcity and domain variability, using attention mechanisms and prototypical networks.

Main Methods:

• Modified a state-of-the-art Cross-Domain Few-shot Learning (CDFSL) approach.
• Integrated attention mechanisms for feature extraction and prototype generation, focusing on salient image regions.
• Utilized prototypical networks for learning category prototypes and classifying new instances.

Main Results:

• Achieved high classification accuracy: up to 96.95% on same-domain and 94.07% on cross-domain datasets.
• Demonstrated stable transfer capability across datasets through result visualization.
• Showcased high visual correlation between model predictions and plant/disease biological characteristics.

Conclusions:

• The modified CDFSL method effectively recognizes plant and disease datasets across domains using generic representations.
• The model exhibits stable transferability and broad applicability, especially when training data includes diverse semantic classes.
• This approach offers a promising solution for efficient image recognition in agriculture and forestry management.