Research On Classification Of Fresh Tea Based On Machine Learning

As a major country in tea drinking,China ranks first in the world in terms of tea garden area and annual tea output.With the vigorous promotion of agricultural mechanization,more and more tea companies are beginning to use machinery to pick tea.However,the machine-picked tea mixed with different categories of fresh tea cannot be directly produced and processed,the precise classification of fresh tea leaves has become an important part of the tea production and processing process.Traditional physical-based sorting methods have low accuracy and efficiency,and the existing classification methods based on computer vision cannot accurately classify the single bud,one bud and one leaf,one bud and two leaves,and one bud and three leaves required for production and processing.In order to solve this problem,improve the classification accuracy of fresh tea while reducing the training time of the classification model,combined with the application scenarios of fresh tea classification,two fresh tea classification models are proposed based on traditional machine learning and deep learning respectively.The main research content and results of this thesis are as follows:(1)This thesis proposes a fresh tea classification method based on multiple features and multiple classifiers.Construct a feature vector by extracting 5 geometric features,7 Hu invariant moments and 1 HOG feature of fresh tea images,and using Support Vector Machine to build basic fresh tea classification model.Then,fit the polygon of the fresh tea image and detect special corner points,and calculate the probability matrix and distance matrix,and use distance matrix similarity as the judgment condition to obtain classification results based on shape features.Finally,use KNN to merge the results of the above two methods to obtain the final classification result.This method can better use the shape characteristics of fresh tea leaves to achieve classification.Experiments show that the fusion of classification methods based on special corner points and their distance matrix with classification models such as SVM,KNN,BP neural network can improve the classification effect.And the fresh tea classification method based on multiple features and multiple classifiers is used to classify the fresh tea data set,and the accuracy is 94.2%.(2)This thesis proposes a fresh tea classification method based on transfer learning Inception V3 based on channel attention mechanism.Squeeze and excite the feature map extracted by Inception V3.Through parameter update to learn the feature weights of each channel of the feature map and scale the original features to achieve the effect of suppressing the features with low contribution and enhancing the features with greater contribution,thereby improving the classification effect.Then use the combination of feature extraction and fine-tuning to realize transfer learning.At the same time,the network is hierarchically divided during fine-tuning training,and only part of the parameters of the pre-training model is frozen to transfer learning to reduce the training time.Experiments show that the classification effect of Inception V3 based on the channel attention mechanism is better than that of the original Inception V3.On the fresh tea data set,the recall,accuracy,and F1-Score were increased by 5.0%,4.5%,and4.9%respectively.Using a combination of feature extraction and fine-tuning to transfer learning can reduce the training time while the training effect reaches the fine-tuning method.And the transfer learning method based on layered freezing parameters can adjust the number of parameter freezing without affecting the performance of the model to reduce the transfer learning time.Applying the fresh tea classification method based on transfer learning Inception V3 based on channel attention mechanism to the fresh tea data set for classification,the accuracy is 98.5%.(3)Compare fresh tea classification methods based on traditional machine learning and deep learning,experiments show that under low-performance hardware conditions,the training time of traditional machine learning models only needs 1.88% of the deep learning classification model.Although the classification accuracy is 94.2% which is lower than the 98.5% of the deep learning method,however,limited by the hardware cost of the tea sorting machine,if you want to complete functions such as model retraining and updating locally on the device,traditional machine learning methods have great advantages.And in view of the high-precision classification effect of the deep learning method,it also has an irreplaceable role in some application scenarios.