Enhancing Apple Cultivar Classification Using Multiview Images.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Science

Background:

• Apple cultivar classification presents challenges due to high similarity between classes and variations within a single class.
• Human experts utilize multi-view analysis, considering various apple viewpoints for accurate identification.
• Previous research explored single-view machine learning for apple classification, indicating limitations for complex tasks.

Purpose of the Study:

• To establish and evaluate a multi-view machine learning approach for apple cultivar classification.
• To compare the performance of state-of-the-art multi-view methods against simpler, preprocessed single-view techniques.
• To assess the feasibility of deploying apple classification models on mobile devices with limited resources.

Main Methods:

• Comparison of an ensemble model with two single-network approaches: one trained on all views without specialization, and another using combined views.
• Utilizing a custom apple cultivar dataset for training and evaluation.
• Analysis of model performance based on accuracy, memory footprint, and computational requirements.

Main Results:

• The state-of-the-art ensemble model achieved the highest classification accuracy.
• Combining views into a single image reduced accuracy by 3% but decreased memory requirements by 40%.
• Simpler approaches with enhanced preprocessing demonstrate a viable trade-off between accuracy and resource efficiency.

Conclusions:

• Multi-view classification, particularly ensemble methods, offers superior performance for apple cultivar identification.
• Preprocessed, combined-view approaches present a practical solution for resource-constrained mobile applications.
• Further research can optimize these simpler methods for efficient on-device apple classification.