A multi-criteria evaluation framework for mobile weed classification: Balancing accuracy and efficiency across hardware tiers

Key Points

• To establish a framework for evaluating deep learning models in mobile weed classification while balancing accuracy and efficiency.
• Evaluated deep learning models specifically MobileNetV3-Large and Small for weed classification accuracy and efficiency.
• Developed a multi-tier benchmarking protocol to assess model performance reproducibly.
• Conducted Pareto-frontier analysis to guide hardware-aware architecture selection.
• MobileNetV3-Large and Small models were identified as Pareto-optimal for accuracy and efficiency.
• Lightweight models demonstrated superior performance in terms of latency and memory usage compared to heavier models.
• The multi-tier benchmarking protocol facilitated consistent evaluation across different hardware configurations.