Rapid prediction of green tea quality based on convolutional neural network

The appearance of tea is a core visual indicator of quality evaluation, which directly affects consumer perception and market grading. This study proposed a green tea appearance quality evaluation method based on a convolutional neural network. A green tea image dataset of three shapes was constructed, containing flat-shaped Dafo Longjing tea, sparrow tongue-shaped Huangshan Maofeng tea, and granular-shaped Pingshui Rizhu tea. Dafo Longjing tea and Huangshan Maofeng tea had six grades and Pingshui Rizhu tea had four grades. Approximately 1500 images were collected for each grade of tea samples, and 9608, 9312, and 5768 images were obtained for three kinds of teas, respectively. The AdamX optimizer, ExponentialLR learning rate decay and batch size 32 were used to compare the classification performance of the ResNet18 model and MobileNetV3 model. The results showed that ResNet18 model achieved 98.28%, 99.19%, and 99.74% accuracies in recognizing grade in Dafo Longjing, Huangshan Maofeng, and Pingshui Rizhu. And MobileNetV3 model achieved 98.91%, 99.62%, and 99.39%, respectively. The fluctuation trends of the precision and recall in the two models during grade identification were consistent, reaching more than 98.20% in ResNet18 model and more than 98.90% in MobileNetV3 model. The effect of parameter tuning on performance metrics was similar in three green teas, indicating that these two models were appropriate for usage in various shapes of green tea. The ResNet18 model with the ECA attention mechanism was larger than the MobileNetV3 model, so MobileNetV3 is more suitable for actual quality detection with its lightweight advantage. Both models confirm the effectiveness of deep learning in tea appearance quality evaluation and provide technical support for automated grading.