Tea selection method based on morphology feature parameters

The color between stalks and leaves of tea in summer and autumn is similar, which means the traditional color sorter is difficult to sort based on optical characteristics. To realize the rapid modeling of tea selection algorithm and improve the sorting accuracy, a method for sorting the fine and bad products of tea by multi-feature vectors based on the morphological characteristics was introduced in this paper. First, Wuyishan Dahongpao tea was selected as a test sample to collect images during the dynamic drop process. The blue element image was extracted, and single sample’s binary image and edge were obtained by analysis of whole image connection area. Then, feature extraction program was developed based on image processing algorithm to extract morphological feature parameters of the tea samples automatically. Four simple shape descriptors-the sample perimeter, area, the length and width of minimum bounding rectangle were extracted. On this basis, eight complex shape descriptors-circularity, rectangularity, linearity, slightness, diameter, diagonal of minimum bounding rectangle, compactness and centroid were calculated. In addition, the random forest algorithm was used to determine the above features weight, the feature was selected according to weight threshold. Finally, logistic regression (LR), decision tree (DT) and support vector machine (SVM) that three different classification algorithms were established to classify the samples, verify the validity of the features and analyze the effects of different classification algorithms on the classification of tea. The original data were normalized and randomly segmented 80% used for training, 20% for testing. 10-fold cross-validation was used to select the optimal parameters of the classification model, and the training dataset was randomly divided into 10 parts, of which 9 parts were used for training, and the remaining 1 part was used for verification. According to the above machine learning system parameter optimization process to obtain the logical regression, decision tree and support vector machine optimal model, and statistical the final evaluation results on test dataset. The test results showed that: 1) The circularity weight was the highest, at 0.467, and five eigenvectors of circularity, rectangularity, linearity, perimeter and compactness were finally selected with the weight threshold value which was 0.05; 2) In the test dataset, the average accuracy F1 of the three classification algorithms was 0.924, suggesting that the established tea morphological feature descriptors has certain separability and better effect; 3)When testing test-dataset, the accuracy score was 91.7% and F1 score of logistic regression (LR) was 92.9%, the accuracy score was 91.7% and F1 score of support vector machine (SVM) was 94.7%.Support vector machine (SVM) algorithm was the best recognition effect in three classification algorithms; 4) From three different classification algorithms assessment score deviation, we can see that the generalization ability of the logic regression algorithm was stronger, the decision tree algorithm has a greater risk of over fitting. We get the lowest accuracy and F1 score of the logistic regression algorithm, while the support vector machine accuracy and F1 score were the highest, so in the evaluation of eigenvector comparability, multiple algorithms can be selected to evaluate the results of the average as the final basis for evaluation. In the experiment, we acquired dynamic image, which stay in line with the actual working conditions of the tea selection process, and can be extended to the actual processing of tea production.