Study on the Tea Pest Classification Model Using a Convolutional and Embedded Iterative Region of Interest Encoding...

Area of Science:

• Agricultural technology
• Computer vision
• Plant pathology

Background:

• Tea diseases significantly reduce crop yield, necessitating effective management strategies.
• Current computer vision methods struggle with classifying tea diseases due to lesion variability, symptom overlap, and complex backgrounds.

Purpose of the Study:

• To develop an advanced IterationVIT model for accurate tea disease classification and diagnosis.
• To improve the efficiency and robustness of automated tea disease identification systems.

Main Methods:

• Proposed IterationVIT model integrating convolutional layers for local feature extraction and an iterative transformer for global feature analysis.
• Utilized a dataset of 3544 tea leaf images encompassing eight categories (seven diseases and healthy leaves).
• Incorporated attention mechanisms and bilinear interpolation for precise disease localization.

Main Results:

• IterationVIT achieved 98% classification accuracy and a 96.5% F1 score with a patch size of 16.
• Demonstrated robustness with over 80% accuracy on augmented images (blurred, noisy, highlighted).
• Outperformed mainstream models in accuracy (8% higher) with reduced training samples and showed comparable generalizability across public datasets.

Conclusions:

• The IterationVIT model offers superior performance for tea disease identification compared to existing methods.
• The model's ability to accurately locate disease regions, confirmed by CAM visualization, highlights its practical diagnostic potential.
• IterationVIT shows promise for enhancing tea cultivation through effective, automated disease management.