Nondestructive Apple Internal Quality Estimation Using Dielectric Signal Processing

Currently among all apples of our country, only2%apples can up to the exportstandards. The fruit internal quality is finally determined by the contents of physicochemicalindicators, but in order to test the contents of physicochemical indicators, we not only needthe professional testing institutions to finish this laborious work, but also must destruct theapple samples to get the test data from the fruit juice. With the further research on therelations between dielectric features and physicochemical indicators, it has beendemonstrated that fruit dielectric features have a very close relationship with their internalquality. Our research analyzes the integrated linear correlation between apples’ dielectricfeatures and their physicochemical indicators firstly, and then employs several apples’dielectric features to detect apples’ quality. Combining with the feature selection andclassifier methods, we innovatively propose a framework to detect the fruits integratedquality with their dielectric features. Our main research contents and conclusions are asfollows:(1) Integrated linear correlation between apples’ dielectric features and theirphysicochemical features was analyzed using methods of canonical correlation analysis andsparse canonical correlation analysis. The dielectric features and physicochemical featurescan be regarded as multivariable systems, and the integrated multivariate statistical analysiscan be achieved by extracting their respective canonical variables. Finally the experimentalresult shows that the integrated linear correlation coefficient is0.793, and it indicates therelationship between apples’ dielectric features and their physicochemical features is close tothe highly relevant.(2) Several feature selection algorithms were employed to select the best apples’dielectric features. In the previous research, dielectric features are chosen manually and havenot conducted the dielectric feature selection based on their research purpose. In our research,the number of dielectric features is108. In order to select the key features for the predictionof the apples’ quality, and to reduce time and space complexity of the detection model, werealized several feature selection algorithms. Combining with three classifier models, we also compared and analyzed their experimental performance in the process of apple qualityprediction.(3) An apple quality detection model based on their dielectric feature selection wasconstructed. In our experiments, the quality of apples is divided into five classes beforehand.After getting the apples’ dielectric features, we analyzed their low rank character, and thenanalyzed the experimental performance of the sparse representation classifier. Meanwhile, wealso compared and analyzed the experimental performance of some representative classifiermodels in pattern recognition fields, i.e. support vector machine classifier and artificial neuralnetwork classifier. The final experimental results show that, the model constructed by thecombination of feature selection method of greedy selector and the classifier method ofsupport vector machine achieves the best performance. The detection accuracy of applequality is98.3%when6dielectric features are selected.