Intelligent pear variety classification models based on Bayesian optimization for deep learning and its...

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate pear variety classification is vital for agricultural efficiency and consumer satisfaction.
• Traditional methods often struggle with large datasets and image variations, including noise.
• Deep learning offers potential but requires careful hyperparameter optimization.

Purpose of the Study:

• To apply Bayesian optimized deep learning for accurate classification of nine pear varieties.
• To evaluate the impact of dataset configuration and noise on classification performance.
• To enhance the transparency and reliability of deep learning models in agriculture through interpretability techniques.

Main Methods:

• Utilized Bayesian optimization (BO) to automatically tune hyperparameters for deep learning models.
• Trained and evaluated models on two challenging datasets with varying Gaussian white noise intensities.
• Employed feature visualization, strongest activations, and LIME for model interpretability.

Main Results:

• Achieved high classification accuracies: 97.29% on dataset A and 90.39% on dataset B.
• Demonstrated that dataset configuration significantly influences classification outcomes.
• Attained 100% accuracy on the Fruit360 dataset with optimal BO hyperparameter tuning.

Conclusions:

• Bayesian optimized deep learning effectively addresses hyperparameter optimization challenges in agricultural CNN applications.
• Dataset configuration plays a critical role in the performance of pear classification models.
• Interpretability methods increase the trustworthiness and applicability of deep learning in agriculture, paving the way for broader adoption.