High precision banana variety identification using vision transformer based feature extraction and support vector...

Area of Science:

• Agricultural science
• Computer vision
• Machine learning

Background:

• Bananas are globally significant fruits, valued for nutrition and flavor.
• Deep learning (DL) has advanced agricultural diagnostics, but banana variety classification remains challenging, especially at early stages.
• Identifying subtle features for accurate classification requires sophisticated methods.

Purpose of the Study:

• To develop a novel hybrid framework for accurate banana variety classification.
• To integrate Vision Transformer (ViT) for global semantic features with Support Vector Machines (SVM) for robust classification.
• To address challenges in identifying subtle features and data imbalance in banana classification.

Main Methods:

• A hybrid framework combining Vision Transformer (ViT) and Support Vector Machines (SVM) was developed.
• The framework was evaluated on two datasets: BananaImageBD (four-class) and BananaSet (six-class).
• Self-supervised and semi-supervised learning mechanisms were employed within the ViT model.

Main Results:

• The hybrid framework achieved high classification accuracy rates: 99.86% on BananaSet and 99.70% on BananaImageBD.
• Performance surpassed traditional methods by 1.77%.
• The ViT model effectively extracted nuanced features crucial for agricultural applications.

Conclusions:

• The proposed hybrid DL framework establishes a new benchmark for automated banana variety detection and classification.
• ViT's ability to extract subtle features is critical for agricultural diagnostics.
• This approach shows significant potential for advancing precision agriculture and automated fruit classification.