Classification of tea quality grades based on hyperspectral imaging spatial information and optimization models

Abstract

To achieve efficient classification of tea quality grades, spatial information from hyperspectral imaging (HSI) technology is proposed as the research focus. The principal component analysis (PCA) method is employed to extract the first three principal component images of tea spectral images, and the PCDS1 and PCDS3 datasets are constructed based on the individual principal component images and the combination of all three principal component images, respectively. Discriminative models for tea quality grades are established using ResNet-50. Without the use of enhancement strategies, the discriminative models established using the PCDS3 dataset achieve better recognition performance, indicating that the strategy of integrating spatial information features contributes to improving model performance. In the case of small sample sizes, transfer learning strategies and image enhancement strategies are employed to enhance model accuracy. The ResNet-50 model using transfer learning strategies exhibits superior performance, achieving a recognition accuracy of 86.15%. To further improve the model’s performance, the particle swarm optimization (PSO) algorithm is utilized to optimize hyperparameters, resulting in an improved model accuracy of 89.23%. Addressing the issue of the PSO algorithm easily falling into local optima, we propose a Two-Strategy Particle Swarm Optimization (TSPSO) algorithm. Experimental results demonstrate that TSPSO significantly outperforms PSO, enabling the identification of more appropriate hyperparameters. The optimal TSPSO-ResNet-50 model achieves a recognition accuracy of 92.31% on the test set. The modeling strategy that combines image information with optimization models is well-suited for identifying the quality grades of tea.