Abstract: Black tea is the second largest type of tea in China, and its processing involves four stages: withering, rolling, fermentation, and drying. Fermentation plays a crucial role in the processing of black tea, closely related to its color and taste. Mild fermentation or excessive fermentation can affect the quality of black tea, thereby affecting its market value. In addition, black tea with moderate fermentation is more stable and less prone to spoilage during storage. Mild fermentation or excessive fermentation can lead to a decrease in the preservation of tea, making them susceptible to external factors and losing the unique quality of black tea. Therefore, the importance of precise control over the fermentation quality of black tea is self-evident. At present, the control of black tea fermentation quality mainly relies on the experience of tea makers in "observing color, smelling aroma, and touching texture", which is highly arbitrary and subjective, resulting in inconsistent black tea quality and inability to achieve standardization in production. In addition, considering the many limitations faced by the deployment of the model in actual production environments, this paper adopts machine vision technology to propose an improved deep learning model that accurately judges the fermentation quality of black tea based on image features. Firstly, we conducted experimental comparisons on the seven selected convolutional neural network models under the same conditions, and all of them have the ability to distinguish the quality of black tea fermentation. Among them, the Ghostnet model showed the best discriminative performance, so it was chosen as the teacher model. Taking into account both the discriminative performance and complexity of the model, Mobilenetv3_small will be used as the student model after comparison. Secondly, taking AdamW, SGD, and RMSProp as the research objects, a comparative experiment was conducted to compare the discriminative performance. Based on this, the selected student model and teacher model were replaced with RMSProp optimizer, which improved the discriminative performance of the model without increasing its complexity or changing its speed. Afterwards, the loss functions of the student model Mobilenetv3_small and the teacher model Ghostnet were simultaneously changed to CE Loss. Without changing the complexity of the models, their discriminative performance was further improved. On this basis, we set the range of Kd_ratio between 0.1 and 2.0, and used the SoftTarget method to conduct knowledge distillation experiments on the student model using the teacher model. The results showed that when Kd_ratio was 0.4, 1.7, and 1.9, the model's discrimination performance was significantly improved while maintaining complexity and speed. In contrast, when Kd_ratio was 1.9, the discriminative performance of the model was improved the most. The improved model achieved Accuracy, Precision, Recall, and F1 of 96.93%, 95.15%, 95.79%, and 95.46%, respectively, for black tea fermentation quality discrimination without increasing model complexity or changing model speed. Compared with the original model, these scores increased by 0.94, 1.91, 1.11 and 1.52 percentage points respectively. This study achieved precise control of the fermentation quality of black tea, meeting the requirements for judging the fermentation quality of black tea and providing strong technical support for the digital and intelligent processing of black tea.