Tea leaf classification method

Abstract

The invention discloses a tea leaf classification method combining a ghost attention capsule network and knowledge distillation. The method comprises the following steps: collecting tea leaf images; data enhancement and data set establishment; constructing a ghost attention bottleneck layer and a tea leaf classification model; pre-training the ResNet50 model and obtaining the weight of the ResNet50 model; training a tea leaf classification model in a growth knowledge distillation mode; and verifying the performance of the tea leaf classification model. According to the method, the requirement for the number of training samples is reduced, and the small-scale data set processing capacity of the model is enhanced; the model can adaptively adjust a weight coefficient so as to emphasize important feature channels and weaken irrelevant feature channels at the same time; while the huge parameter matrix of the teacher model is transmitted to enhance the comprehensive performance of the student model, the purpose of adaptively reducing the dependency degree of the student model along with distillation is achieved.

Description

technical field [0001] The invention relates to the technical field of tea green classification, in particular to a tea green classification method combining ghost attention capsule network and knowledge distillation. Background technique [0002] When picking tea leaves, tea greens of different grades will be mixed together, and the traditional method of sorting only by the experience of tea makers will cause the classification results to be greatly affected by human subjective factors, which will seriously restrict the development of tea production enterprises. Efficiency increases time and production costs, which is not conducive to large-scale batch customized production and transformation to the famous tea industry. [0003] Although deep learning technology has achieved excellent performance in classification problems in many fields, due to the limitations of regional factors in tea planting and drinking, research on tea green classification based on this technology is...

AI Technical Summary

Problems solved by technology

[0005] (1) The traditional model based on the convolutional layer will require a large number of samples in order to train the weight matrix. In practical applications, the collection of brown and green images faces many challenges, and the lack of training samples will restrict the classification performance of the model.

[0006] (2) The existing model cannot measure the importance and degree of correlation of different channels in the matrix features of dark brown images, and adaptively adjust the weight parameters of each channel to achieve the purpose of strengthening effective features and suppressing irrelevant features

[0007] (3) The existing technology mainly uses a single model for training, and the collaboration between multiple models has not yet been realized, and it is difficult to further enhance the performance of the dark green classification model with the help of rich weights of complex models

Images