What does this research mean for the field?

A lightweight CNN model (AT-ShuffleNet) optimized by knowledge distillation can accurately and efficiently classify the fermentation stages of black tea with 91.2% accuracy, overcoming the limitations of manual determination. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

The study aims to develop a lightweight model for accurate classification of different fermentation stages in black tea production.

May 20, 2026Open Access

Classification of Different Fermentation Stages of Black Tea Using a Lightweight CNN Optimized by Knowledge Distillation

Key Points

The study aims to develop a lightweight model for accurate classification of different fermentation stages in black tea production.
Developed a lightweight CNN model called AT-ShuffleNet using knowledge distillation.
Collected images of Fuding and Tieguanyin tea at various fermentation stages.
Optimized models using focal loss and poly loss, testing multiple distillation methods and ratios.
The AT-ShuffleNet achieved a precision of 89.11% and recall of 90.16%.
Kappa statistic was 89.29%, with an accuracy of 91.2% and F1 score of 89.53%.
Successfully deployed the model on edge devices for real-time classification in tea processing.

Abstract

In red tea production, fermentation is critical for flavor. However, manual determination of its stages is inaccurate and inefficient, often spoiling flavor and lowering product value. To solve this, this study combines CNN and knowledge distillation to build a lightweight classification model, AT-ShuffleNet, for accurate, efficient stage identification in real processing. It collected images of Fuding and Tieguanyin tea at different fermentation stages. ResNet (teacher model) and ShuffleNet v2-0.5 (student model) were used for distillation. Focal and Poly Losses optimized both models to tap distillation potential. STD, MGD, SPKD, ATD, and KD methods were tested at various ratios to find the optimal strategy, forming AT-ShuffleNet. The lightweight model performed well: P (89.11%), R (90.16%), Kappa (89.29%), ACC (91.2%), F1 (89.53%). It addresses manual limitations, enabling accurate classification and reducing deployment issues in unstructured environments. For industrial validation, it was deployed on edge devices and integrated into a self-developed WeChat mini-program.

Read Full Paperexternally

Mark Helpful

Bookmark

Relay

View Full Paper