Research On Tea Classification Algorithm And Sorting Equipment Design And Experiment

Famous tea has been mainly picked by hand,which has a large demand for labor and low picking efficiency.Mechanized picking can improve the picking efficiency of fresh leaves,but fresh tea leaves plucked by machine are often mixed and have a high broken rate and only suitable for making general tea.Therefore,the machine-picked tea is sorted according to a certain grade standard based on machine vision to obtain high-quality tea,which is an effective way to solve the problem of picking famous tea.In this paper,five different types of tea are selected and marked according to grade,and the image data set is made.Two kinds of feature libraries are obtained based on manual extraction of morphological,texture and color features and automatic feature extraction based on convolution neural network,and the classification and recognition models are established and compared respectively.Finally,the algorithm model based on deep convolution neural network is selected,and the multi-target location and recognition of tea materials are realized by combining region segmentation and convolution neural network.On this basis,the tea sorting equipment is developed and tested.The main research work and achievements are summarized as follows:1)Three tea sample image data sets were made to analyze the influencing factors in the process of tea classification recognition,and two tea image data sets were used to analyze the influencing factors in the process of inferior tea products recognition,and the categories were marked according to the grade.2)42 features of sample shape,texture and color were extracted,and a method for fast selection of effective features under multiple eigenvectors is analyzed.The effects of logistic regression(LR),decision tree(DT)and support vector machine(SVM)on the recognition of five kinds of tea samples were explored,and the best recognition models of five kinds of tea samples were obtained.Through the analysis of the results of the best recognition models of five kinds of tea samples,it was found that the tea processing technology would have an impact on the accuracy of tea classification.3)The method of automatic tea feature extraction and recognition based on convolution neural network is proposed,and the automatic extraction of complex features of tea samples is realized.The effects of batch,learning rate parameters and sample number on model training are explored,and the two model over-fitting optimization methods of L2 regularization and Dropout are compared.Then,the optimal recognition models of five kinds of samples are obtained and tested.The experimental results show that convolutionneural network can not only extract features automatically in the process of tea sample classification,but also achieve higher recognition accuracy than the recognition model based on manual extraction of morphological,texture and color features.4)The method of target location and recognition based on region segmentation and convolution neural network is proposed,and the fast segmentation and recognition of multi-target image is realized.The image segmentation and recognition of more than 10 targets can be completed in 34 ms,which can effectively solve the problem of target location based on convolution neural network image classification and improve the speed of multi-target image recognition.5)The convolution neural network feature extraction and region segmentation are applied to the actual sorting platform,and the tea online sorting method based on deep learning algorithm is realized.6)In order to solve the feeding problem of fresh leaves and raw tea,two different feeding mechanisms of vibrating feeding mechanism and negative pressure rotary blanking mechanism are designed to realize the decentralized feeding of fresh leaves and raw tea.Two different executive mechanisms of air blowing and suction are established,and the matching control system is built,and finally the dynamic sorting of materials is realized.7)The experimental results of sorting equipment show that the multi-target recognition model based on region segmentation and convolution neural network is stable and the recognition rate is more than 95%.The average sorting rate of fresh leaves was 90.5% and92.8% respectively,which decreased compared with the recognition results.It was found that the air suction structure was not suitable for the classification of fresh leaves,and the tea varieties would affect the results of fresh leaves classification to a certain extent.In the raw tea sorting experiment,the sorting rate of the air suction actuator reached 98.5%,and the misselection rate was only 1.5%,and the sorting rate of the air blowing actuator was97.2%,which achieved a better separation effect.At present,the work efficiency of single-channel sorting of fresh leaves can be equivalent to about four manual workers,and the efficiency will be further improved with the continuous optimization of hardware.