Research On Internal Quality Of Apple Multi Lable Classification Method Based On Random Forest

As one of the four major fruits,apple occupies a very important position in human daily life.The apple production in China ranks first in the world,but its export volume only accounts for 1.46% of the total output.An important factor that hinders the export of fresh fruit in China is apple’s sorting,testing capabilities,and inspection speeds,which do not meet market demand.With the rapid development of social science and technology,evaluating the quality of apples is no longer limited to the evaluation of external indicators such as color and size.People are paying more and more attention to the nutritional value and internal quality.Advanced fruit sorting requires and detection capability based on the detection of fruit internal quality.On the basis of internal quality testing,the internal quality of fruit is determined by its internal sugar content,moisture content and other physical and chemical indicators.Therefore,it is of great significance to seek a method for detecting the internal quality of apple sugar,hardness,moisture content and the like for the graded sales and export of apples.A large number of studies have shown that the dielectric characteristics and the internal indicators of fruit have a very close relationship,and the measurement of dielectric characteristics can be done without loss.This study infers the physical and chemical characteristics of apples based on the dielectric characteristics.The main contents of this study are as follows:(1)Build an Apple internal quality classification model based on random forest.Based on the working principle of the random forest classifier,the selection and function of its main parameters were analyzed.There are 108 kinds of intermediate electrical characteristics in this experiment,and there are 8 kinds of physical and chemical characteristic labels.The physical and chemical characteristics are divided into 5 levels,a total of 8 × 5 label information.Divide 500 apples into 10 subsets,use 10 subsets,use one of the 10 subsets as the test set,and use the remaining 9 as training sets to train multiple decision trees to form a random forest using the 10-fold cross validation method..The training results show that the random forest classifier can effectively deal with the apple internal quality multi-label classification problem,and the larger the forest scale is,the higher the classification accuracy is.The random forest using the information gain objective function has better classification performance than other random forests..Compared with the SVM algorithm,random forests have better classification accuracy..(2)Select the output category of the random forest.Different from the traditional decision tree method,this experiment treats the random forest as a whole,so that the fusion information is separately included in each leaf node,each leaf node corresponds to an apple,and the apple is labeled with physical and chemical characteristics.When the category is selected,the TFIDF algorithm returns the output category of the random forest.The experimental results show that the random forest output categories returned by the TF-IDF method can effectively correlate dielectric characteristics with related physical and chemical characteristics..(3)To sort the output category.In this study,apple physicochemical characteristics of 8 indicators were assigned into 5 grades.Because each decision tree of the random forest is independent,after labeling apple labels in the label prediction probability,this study sorts the number of output categories,and systematically allocates Apple’s physicochemical characteristics of the best labeling apple.In this experiment,the Rank SVM algorithm was used to sort,and three frame models were applied to Rank SVM.They were f(ci)=ci,f(ci)=ci2 and f(ci).The proposed system selects the frame model of f(ci).The experimental results show that the effect of f(ci) on apple quality classification is better through the selection of the system,and the distribution of the label level is more consistent in the classification prediction,and the root mean square error is 0.51.