Abstract: Considering that traditional sensory evaluation is highly subjective and inefficient, this study designed a model for the detection of tea appearance quality using a residual network optimized by feature aggregation and attention mechanism. A total of 1 800 tea sample images covering 22 grades of three types of tea (Keemun black tea, Huangshan Maofeng, and West Anhui yellow tea) were collected and divided into training and test sets at a ratio of 8:2. Based on the residual network with 50 layers (ResNet50), transfer learning was applied to retain general visual feature extraction capabilities, the top fully connected layer was replaced to adapt to the classification task, and a convolutional block attention module (CBAM) and a path aggregation network (PANet) were embedded. A multi-task architecture was designed and optimized using a joint loss function to achieve recognition of tea types and grades. The classification accuracy of the proposed model reached 100% for tea types and exceeded 93% for tea grades, demonstrating good classification performance. This model is expected to provide technical support for tea appearance recognition and contribute to the development of intelligent tea quality control technology.

Key words: deep learning; residual network with 50 layers; tea; multiple types; recognition; convolutional block attention module; path aggregation network