Adaptive classification and grading model of cigar wrapper leaf based on improved ResNet algorithm

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Cigar wrapper leaf grading is crucial for quality but relies on inefficient manual sorting.
• Current manual methods suffer from low efficiency and inconsistent grading standards.
• A need exists for an intelligent, automated system to improve cigar wrapper leaf classification.

Purpose of the Study:

• To develop and validate a deep learning framework for automated cigar wrapper leaf grading.
• To enhance feature extraction using a Mask-augmented Fourth Channel (Mask4Ch) module with segmentation masks.
• To improve classification accuracy and standardization in the cigar industry.

Main Methods:

• A ResNet-50 backbone integrated with a Mask4Ch module was used for feature extraction.
• A dual-head training strategy combined Cross-Entropy (CE) and Cumulative Ordinal Regression (CORAL).
• Optimizations included weighted random sampling (WRS) and exponential moving average (EMA) for improved generalization.

Main Results:

• The model achieved 94.39% accuracy, 0.950 macro-F1 score, and 0.964 weighted Kappa (QWK).
• Mean Average Precision (mAP) reached 0.985 on the test dataset.
• The results demonstrate the effectiveness of deep learning for automated cigar wrapper grading.

Conclusions:

• Deep learning offers a powerful solution for automated cigar wrapper leaf grading, surpassing manual limitations.
• The developed framework provides a technical foundation for optimizing grading systems and potential mobile deployment.
• This research advances automation and standardization within the cigar production industry.