Research On Apple Grading Based On Multi-feature Ensemble Learning

Apple is a common fruit in daily life.The scale of apple production and planting in China has reached more than 50% of the global scale.Apple grading is a key step after apple picking and before sales.The grading technology of apple in China is still based on manual and mechanical grading,which consumes a lot of labor costs.However,the grading results are quite different,which cannot meet the grading requirements of the international market and seriously lacks market competitiveness.In this thesis,Red Fuji apple is taken as the research object,and the apple grading method is studied based on apple multi-features and ensemble learning.The main work is as follows:(1)Aiming at the problem that apple sample images contain irrelevant information to apple grading,the bimodal threshold method,median filtering method and morphological closed operation combined with canny operator are used to segment and smooth the background,noise and edge subtle points of apple images,so that the processed apple images meet the grading requirements.(2)Extract the color,texture,shape,size and other external features of apple.Apple color characteristics are obtained by color quantification and red coloring rate.Gray level cooccurrence matrix is used to extract texture features.The minimum and maximum radius ratio method and circularity method are used to describe the shape of apple.The minimum circumcircle method is used to calculate the apple diameter.This method can complete the extraction of apple multi-features and provide a basis for subsequent apple grading.(3)The Mobile Net V3 network model is built using the Pytorch framework,and 1200 apple sample images are divided into training set,verification set and test set according to the ratio of 7:2:1.Aiming at the problem that the Mobile Net V3 network model classifier is too single and the classification accuracy is too low,a parallel integrated multi-classifier based on Logistic and SVM is designed in combination with the integrated learning theory to improve the full connection layer and classification layer in the Mobile Net V3 model.Comparative experiments show that the classification accuracy of Mobile Net V3-SL network model is 96.62%,which is higher than that of Mobile Net V3-S and Mobile Net V3-L single classifier network models.(4)Aiming at the problem that the SE module in the Mobile Net V3-SL network model has the problem of missing feature information in feature extraction,the BAM and CA modules are used to improve the attention mechanism on the basis of adding the integrated classifier.The comparative experiment proves that the classification accuracy of the apple data set on the BAM-Mobile Net V3-SL and CA-Mobile Net V3-SL network models is 96.93%and 97.11% respectively.Considering the training time and model size,the CAMobile Net V3-SL network model has a better effect in apple grading.(5)The original Mobile Net V3 network model is transplanted into the K210 development board.Through the grading experiment of apples,the accuracy of grading is 93.75%.The accuracy of apple grading using CA-Mobile Net V3-SL model on PC reached 96.25%,which verified the feasibility of CA-Mobile Net V3-SL network model in apple grading algorithm.