Abstract: Longjing tea appearance features are crucial for assessing tea quality. In order to intelligently evaluate the appearance characteristics of Longjing tea and quickly recognize the grades, a method combining image features and intelligent algorithms for appearance quality assessment was proposed. Images of the six grades of the standard Dafo Longjing tea were collected and their shape, color, and texture features were quantified and extracted to construct a database of tea appearance grade characteristics. The appearance features were fused and input into five machine learning models, namely, Decision Tree, Random Forest, K-Nearest Neighbor, Gaussian Bayesian, and Support Vector Machine, for grade recognition training. Three convolutional neural networks, VGG16, ResNet18 and DenseNet121 were compared to build a deep learning model for Longjing tea appearance evaluation. Then the model was optimized by optimizers and learning rate decay. The results show that the fusion of shape and texture data combined with Support Vector Machine was the optimal model for grade recognition, with an accuracy of 91.14% and an F₁ score of 91.20%. The ResNet18 network structure was chosen to establish the optimal model for appearance evaluation. After optimization by Adadelta optimizer and CosineAnnealingLR learning rate decay, the recognition accuracies of Longjing tea’s flatness, straightness, tenderness, and color all improved, reaching 99.21%, 99.51%, 99.56%, and 99.68%, respectively. This study provided a theoretical foundation for the digital evaluation of the appearance quality of tea.

Key words: Longjing tea, image recognition, grade, appearance, machine learning