Background: Accurate tea blending assessment and sample matching are critical in the tea production process. Traditional methods face efficiency and accuracy challenges, which can be addressed by advances in computer vision and deep learning. This study developed an efficient and non-destructive method for fast tea grading classification, blending ratio evaluation, and sample matching. The method trained a Residual Network (ResNet) model on an enhanced dataset of tea images and used Convolutional Block Attention Module (CBAM) to improve the model's feature-extraction ability.
Results: The enhanced grade classification model achieved 95.05% accuracy for oolong tea and 99.13% accuracy for black tea, outperforming other deep-learning models such as EfficientNet, MobileNet, and VGG16. For oolong tea blends, the model demonstrated greater efficiency than manual evaluation with an average absolute error of 2.26%. In black tea sample matching, the model achieved an average error of 3.34%.
Conclusion: These results highlight the importance of attention mechanisms in improving the analysis of images with intricate textures. The integration of deep learning and attention modules enhanced the accuracy and efficiency of tea quality evaluation processes effectively. This study underscores the transformative potential of intelligent classification and analysis methods in modernizing tea production, ensuring higher standards of consistency and quality. © 2024 Society of Chemical Industry.
Keywords: attention mechanism; computer vision; deep learning; sample matching; tea blending; tea grade classification.
© 2024 Society of Chemical Industry.