| Tea is a traditional economic crop in China,as a major tea-drinking country,our country’s tea garden area and annual tea output ranks first in the world.At present,most of the steps of tea processing,such as killing,kneading,barring and drying,are done by manual labor.In order to realize the intelligence of tea production and improve the quality of tea,an automatic tea processing line can be built in the future,which only needs to pour the tea leaves acquired from the tea farmers into the designated location,and can produce tea products that can be sold directly.When processing fresh leaves,the first step to do is the grading of fresh leaves,a good grading system is beneficial to improve the quality of tea leaves.In this dissertation,the fresh tea leaves of four grades of shrub-type,small-leaf tea species widely distributed in north and south of Yangtze River were taken as samples,namely:bud head,one bud and one leaf,one bud and two leaves,and bulk tea.Most of the existing tea classification methods are based on machine learning,which have high dependence on the extraction of features and high requirements on equipment.To address the above problems,this dissertation will propose several lightweight neural networks,which can both automatically extract sample features and reduce the computational volume and number of parameters.Based on several classical lightweight networks to train and test the tea samples,the respective model results are obtained,and the most suitable model is determined after the comparative analysis of the results.The main model selected in this dissertation is GhostNet model,which can "cheapen" the redundant feature maps in training,using fewer parameters to generate more feature maps,reducing the computational cost of the general convolutional layer while maintaining the recognition performance,and has achieved The results are excellent for Image Net classification.Accordingly,the main work of this dissertation is as follows:(1)Pretreatment of fresh tea leaf sample data.First,2085 original tea leaf images were divided into training set,validation set and test set according to 8:1:1,and then 128*128 pixel size data samples were intercepted from 2085 original tea leaf images of 2592*1944 pixel size.In order to ensure the validity of the intercepted samples,the OTSU method was used to perform binary segmentation for the grayscale map,and then to judge whether the number of pixels in the white tea leaf part accounted for more than 60% of the overall total pixels,and the intercepted tea leaf fresh leaf samples were named according to the rules to facilitate the subsequent operation.With this criterion,10 samples were intercepted in each original tea leaf image to obtain 20850 tea leaf samples.(2)The processed data were put into Shuffle Net model,Mobile Net model and GhostNet model for training,and the curves of their loss values and Top1 values were analyzed,and all three models had better performance results.Among them,the GhostNet model stabilizes the fastest during the training process and has the best value after stabilization.Then the test results of the three models are analyzed,their accuracy and test time are compared,and their confusion matrix is drawn.The Top1 values of the three models are 98.269%,98.412% and 98.558%,respectively.After a series of comparative analyses,GhostNet model has a moderate number of parameters and computation,the highest Top1 value,and the model has a suitable testing time.It can be concluded that GhostNet model is the most suitable model in this tea fresh leaf grading study.(3)After determining the GhostNet model as the best model,the prediction of tea leaflets and voting to obtain tea big picture grades were performed based on the GhostNet model.In 208 test sets,206 were classified correctly,with an accuracy rate of 99.04%.The reduction from small picture voting to large picture increases the practical applicability of this model,making it more applicable in the intelligent and non-destructive tea processing process of grading. |
PDF Full Text Request