PD-ViCo: an explainable AI-based contrastive captioner vision transformer with patch dropout for multi-class brinjal...

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Brinjal (eggplant) production in South Asia is significantly impacted by diseases, affecting yield and quality.
• Manual disease diagnosis is labor-intensive, subjective, and error-prone, highlighting the need for automated solutions.
• Existing automated methods often struggle with data imbalance and generalization in real-world agricultural settings.

Purpose of the Study:

• To develop and evaluate PD-ViCo, a novel, lightweight transformer-based model for accurate brinjal fruit disease classification.
• To address challenges of data imbalance and improve model robustness using advanced techniques.
• To provide an interpretable and practical tool for agricultural disease diagnosis.

Main Methods:

• Developed a new dataset of 1,823 brinjal images covering five disease classes and healthy samples.
• Employed a Simple Vision Transformer (ViT) enhanced with Patch Dropout and Contrastive Captioner (CoCa) methods.
• Implemented extensive data preprocessing, class balancing (under-sampling/oversampling), and data augmentation.

Main Results:

• PD-ViCo achieved 99.12% classification accuracy and 97.76% F1-score on imbalanced datasets.
• The model demonstrated superior performance compared to standard ViT and Swin Transformer.
• Explainability methods (Grad-CAM) confirmed model focus on relevant disease-affected regions.

Conclusions:

• PD-ViCo is a highly accurate, interpretable, and efficient model for multi-class brinjal disease diagnosis.
• The model's robustness and generalization capabilities are enhanced by patch dropout and CoCa-style aggregation.
• This research offers a valuable dataset and a practical decision-making protocol for agricultural stakeholders.