结合改进胶囊网络与知识蒸馏的茶青分类方法研究

不同等级茶青的准确分类,对名优茶产业发展至关重要,采用传统感官审评方法进行分选会使结果存在一定的主观性。采集茶青图像建立数据集,结合幽灵注意力瓶颈层与胶囊网络提出一种新型网络模型：GA-CapsNet。通过基于线性衰减比例系数的成长知识蒸馏方法对该模型进行训练,在迁移教师模型参数矩阵的同时,使学生模型随着迭代自适应降低依赖程度。试验结果表明,对比其他同类算法,所提出的方法在小规模数据集上分类性能优异,精确率、召回率及F1-score分别为94.97%、95.51%、95.24%。本研究基于机器视觉与深度学习技术构建了一种GA-CapsNet模型,为解决茶青分类问题提供了一种新思路。

Research on the Classification Method of Tea Buds Combining Improved Capsule Network and Knowledge Distillation

The accurate classification of different grades of tea buds is very important for the development of the famous tea industry. The use of traditional sensory evaluation methods for sorting makes the results subjective. In this research, a data set was established after tea leaf images were collected, and a new network model, GA-CapsNet, was proposed by combining the ghost attention bottleneck and capsule network. The model was trained by the method of growing knowledge distillation based on the linear decay scaling coefficient, while migrating the parameter matrix of teacher model, the student model was adaptively reduced with iteration. The experimental result shows that, compared with other similar algorithms, the proposed method had excellent classification performance on small-scale data sets. The accuracy, recall and F1-score were 94.97%, 95.51% and 95.24%, respectively. Here, a GA-CapsNet model based on machine vision and deep learning technology was established, which provided a new idea for solving the tea leaf classification problem.