The Bayesian mixture expert recognition model for tobacco leaf curing stages based on feature fusion

Area of Science:

• Agricultural Science
• Computer Science

Background:

• Tobacco leaf curing stages exhibit diverse visual features influenced by origin and environment, complicating precise identification.
• Single features or models struggle to capture the complexity of tobacco leaf visual characteristics during curing.

Purpose of the Study:

• To develop a robust recognition model for tobacco leaf curing stages using feature fusion and ensemble learning.
• To enhance the accuracy and interpretability of automated tobacco leaf stage identification.

Main Methods:

• Utilized deep learning models (ResNet34, MobileNetV2, EfficientNetb0) for feature extraction from a tobacco leaf image dataset.
• Employed various feature fusion techniques (concatenate, scaled, adaptive gated) and optimized models for ensemble learning.
• Applied Bayesian optimization to integrate multiple expert models for final recognition.

Main Results:

• The proposed Bayesian Mixture Expert Recognition Model achieved a recognition accuracy of 93.96% on the test set.
• The ensemble model significantly outperformed individual base models in recognition performance.
• The integrated approach effectively captured and recognized complex dynamic visual features of the tobacco curing process.

Conclusions:

• The developed model offers an efficient and robust solution for recognizing tobacco leaf curing stages.
• Feature fusion and Bayesian optimization enhance system adaptability and interpretability in complex environments.
• This research provides strong support for the intelligent upgrading of the tobacco industry through advanced image recognition.