Discrimination of black tea fermentation degree with integrated attention mechanisms

Abstract

Accurately assessing the fermentation degree of black tea is essential for maintaining quality standards during its processing. However, existing methods primarily depend on human expertise, resulting in significant time expenditure and inconsistent performance. Addressing these limitations, this study proposes the Synergistic Adaptive Fusion Network (SAFNet), a recognition model with attention mechanisms and feature fusion, to accurately identify fermentation degree. To better capture fermentation-related features, the Synergistic Convolution Attention Module (SCAM) is designed to integrate global context with local details to retain essential information. Additionally, the Adaptive Multi-Scale Attention Feature Fusion (AMSAFF) module is developed, enabling the adaptive integration of multi-resolution feature maps from different receptive fields. This mechanism improves the model’s capacity to concentrate on key fermentation-related features and reduce interference from irrelevant background information. Given the difficulty in obtaining black tea fermentation images, data augmentation utilizing the DDPM diffusion framework together with the ESRGAN super-resolution technique is employed to expand the dataset. Experimental results on a self-constructed black tea fermentation dataset demonstrate that SAFNet achieves an accuracy of 87.03%, surpassing ResNet50 by 8.6% in Precision, 8.52% in Recall, and 8.6% in F1-score. These findings reveal the capability of SAFNet for black tea fermentation classification and its potential for intelligent tea processing applications.