Convolutional neural network based on transfer learning for discriminating the fermentation degree of black tea

Black tea is among the most widely consumed tea. The fermentation process is crucial for developing the flavor of black tea. Currently, many producers rely on personal experience to gauge fermentation, which can be inconsistent and subjective. Additionally, large models are impractical for use in production. Based on this, this paper introduces a lightweight convolutional neural network utilizing transfer learning to assess the fermentation level of black tea. Initially, we applied a model-based transfer learning strategy and conducted pre-training weight experiments to compare and select the student model and the teacher model. Next, we modified the loss function with PolyLoss and optimizer with AdamW for the student model. Finally, we performed a knowledge distillation experiment on the student model. Results indicated that the improved model's accuracy, precision, recall, and F1 improved by 0.0415, 0.0215, 0.0902, and 0.0645, respectively. This research offers technical assistance for digital production of black tea.