| Panax notoginseng is one of the most representative traditional Chinese medicines in Yunnan.role.The number of heads of Panax notoginseng refers to the number of main roots of Panax notoginseng per 500g(1 kilogram),which is the main basis for the classification of Panax notoginseng in the market,and the weight of a single Panax notoginseng is the main factor determining its selling price.At present,the grading methods of Panax notoginseng taproots mainly include manual grading and mechanical sorting.Manual grading is easily affected by subjective factors,with high labor intensity,low degree of automation and low efficiency.The sorting adopts the method of weighing,and the grading stability is low.Therefore,rapid and accurate grading of the main roots of Panax notoginseng is of great significance for improving the value of Panax notoginseng.In recent years,machine learning algorithms based on machine vision technology have been widely used in agricultural product quality detection and grading due to their high accuracy and fast speed.In this study,a classification model of Panax notoginseng based on machine vision feature fusion was constructed through traditional machine learning and deep learning,to explore the influence of different models on the classification accuracy of Panax notoginseng,and to build a classification system for Panax notoginseng based on the semantic segmentation model.The system provides some references for the market-oriented and intelligent equipment of Panax notoginseng taproot classification.The main research contents and conclusions of this paper are as follows:1.In this study,a variety of Panax notoginseng taproot grading models based on machine vision image feature fusion were established.(1)Extract 40 kinds of features such as shape,size,color,texture,etc.In the case of the same sample size,through the classification of a single feature by different algorithms,the research results show that the three algorithms perform better in color features,and the accuracy rate of the training set has reached more than75%.The SVM algorithm showed the best results on all single features.Through the research on the classification of mixed features by different algorithms,the research results show that the three classification algorithms perform better than other mixed features in shape size and color mixed features,and SVM performs the best in all mixed features..Comparing the singlefeature and mixed-feature classification results,it is found that both mixed-features show better results than single-features.Therefore,the fusion feature is very helpful for the classification of the main root of Panax notoginseng.(2)Using IRIV,VISSA and SRA three feature selection algorithms to reduce the dimensionality of all features,get 22,21,and 10 best feature variable combinations respectively,analyze and compare BP,ELM,SVM models based on full data features And the IRIV-SVM,VISSA-SVM,and SRA-SVM models based on feature selection,the results can be obtained: the IRIV-SVM model has the best classification results,and the test set accuracy rate reaches 95.370%,indicating that the feature selection algorithm can reduce the number of features.At the same time,the efficiency and accuracy of the model are effectively increased.(3)The GWO,GA and PSO algorithms were introduced to optimize the IRIV-SVM model respectively.The results show that the IRIV-GWO-SVM classification model has the best optimization effect,the test set accuracy rate reaches 98.704%,and the optimization effect is increased by 3.334%.2.This study establishes a deep learning model based on semantic segmentation.(1)Three semantic segmentation frameworks,PSPnet,U-net and Deeplabv3+ are selected as the main root classification models of different numbers of heads.ResNet50,VGG16,and xception are used respectively.The convolutional neural network is used as the feature extraction network,and the average pixel accuracy of the category and the average interaction ratio are used as the evaluation indicators.The results show that the Deeplabv3+ model that extracts the main root features of Panax notoginseng through the xception convolutional neural network has the best comprehensive performance,and the MPA on the test set is the best.,MIoU were 77.98% and88.97%,respectively.Therefore,the Panax notoginseng taproot classification model established by Deeplabv3+ is the best.(2)Using the Deeplabv3+ framework for semantic segmentation in deep learning,and using the improved xception network as the backbone feature extraction network,the improved model Imp-Xce-Deeplabv3+ is compared with U-net,PSPNet,and deeplabv3+.The results show that the improved model achieves In order to achieve better segmentation accuracy,it reaches the highest among the two segmentation evaluation indicators of pixel accuracy and average intersection ratio,which are 85.72% and90.32%,respectively.(3)In order to compare the segmentation effects of Xce-PSPNet,XceUnet,Xce-Deeplabv3+,Imp-Xce-Deeplabv3+ models more intuitively,the segmentation effects of different models are visualized,and it can be seen that the segmentation effect of Imp-Xce-Deeplabv3+ is closest to the original Image tagging results.In conclusion,the ImpXce-Deeplabv3+ model has high segmentation accuracy and strong robustness,and can be used as an automatic segmentation model for the main root of Panax notoginseng of different grades3.A set of dynamic sorting control system of Panax notoginseng taproot based on the semantic segmentation model of Imp-Xce-Deeplabv3+ is studied.(1)The matching table of the speed of the conveyor belt and the time of the target reaching the nozzle was calculated under the condition that the frequency of the motor and the distance of the target reaching the nozzle at each level were different.The test shows that the maximum running speed of the conveyor belt is 1.81m/s,and it takes about 0.22 s to complete one frame of image processing.At least 270 main roots of Panax notoginseng can be separated in one minute.At present,a single channel can replace 5 workers.With the continuous optimization of the later hardware,the efficiency continues to improve.(2)The sorting of the main root of Panax notoginseng was tested under static and dynamic conditions.The results showed that the average recognition rate of the model for the main roots of Panax notoginseng of different grades was 81%,and the average error rate was 19%.It can be seen that the model has the best recognition effect on the first-level Panax notoginseng taproot,followed by the fourth-level Panax notoginseng taproot,and the classification effect on the second-level and third-level Panax notoginseng taproot is slightly worse.visualization.The recognition effect of the model for samples at all levels under dynamic conditions was verified.The results showed that the average sorting rate was 77%,the average false selection rate was 22.025%,and the average sorting rate of the samples of Panax notoginseng at different grades was compared with that of the static recognition test.Compared with the average false recognition rate of the static identification test,the average value of the false selection rate has increased by about 3.025%.During the test,it was observed that the sorting rate of the dynamic sorting test decreased and the false selection rate increased.The main reason is due to the irregular shape of the material and the differences in the training environment of the model.And through python’s Tkinter library,a grading visual interface of Panax notoginseng taproots is created,which is easy to operate,has high stability,and can display the effect and grade of Panax notoginseng taproots in real time.In short,the equipment can basically meet the requirements of the processing plant of Panax notoginseng taproot,and accelerate the intelligentization of the field of Panax notoginseng grading. |
