Nondestructive classification of internal quality of apple based on dielectric feature selection

Abstract: In order to reduce the cost of the application of dielectric signals in nondestructive detection of fruits and crops, it is important to find effective methods to select the key features from all other dielectric features. In this paper, we propose a two stage framework to achieve a low cost effective apple internal quality estimation system. In the first stage, we search a compact discriminative dielectric feature sub set. And in the second stage, based on the dielectric features selected by the first stage, we propose a nondestructive apple internal quality estimation system by evaluating several classifiers. In our experiments, the internal quality of Fuji apples is graded into 5 grades according to a compact set of dielectric features which are selected from the 108 dielectric features obtained from 12 dielectric parameters under 9 frequency points ranging from 158Hz~3.98MHz, and all the dielectric features are measured with HIOKI 3532-50 LCR tester and labeled with a number ranging from 1 to 108. Meanwhile, 100 randomly selected apples of each grade, i.e. a total of 500 apples, are used as the experimental samples, and each apple sample is assigned a 5-grade quality label by its weight loss rate (WLR): the fresh apple is classified as Grade One whose WLR is 0, those with WLR is equal to 5%, 10%, 15%, are labeled as Grade Two, Three, and Four respectively, and the apple with brown stain is grouped into Grade Five. During our whole experiments, 80% samples selected randomly from the dataset are used to train the classifier and the other 20% are used to test the classification accuracy. In the dielectric feature selection stage, greedy feature selection, fast clustering-based feature subset selection (FAST), sparse principal component analysis (SPCA), and attribute ranker method with the attribute evaluator of information gain are employed. With the dielectric feature dataset, FAST can only select a fixed number of discriminative dielectric features, while SPCA, greedy selector, and attribute ranker method can adjust the algorithm parameters to control the number of the key dielectric features. The compact set of dielectric features are the most discriminative for apple internal quality estimation. In the internal quality estimation stage, three classifiers are evaluated. They are sparse representation classification (SRC), artificial neural network (ANN), and support vector machine (SVM). According to the experimental results, FAST only selects four dielectric features and the classification rate is about 80%. SPCA tends to select the dielectric features with the same dielectric parameter, and its classification accuracy compared with the other three classifiers is mediocre; the performance of greedy selector is significantly outstanding. When the classification rate is higher than 90%, the number of the selected features of greedy selector is generally, lower than 10. With the greedy selector, the best classification rates are 91.22% and 95.95% when the number of the selected dielectric features is 4 and 10 respectively. The results show the dielectric features are highly relevant to the apple internal quality, and apple internal quality can be estimated with a compact set of dielectric features. The experimental results provide a reference for quick and nondestructive detection of the quality and insect pests to fruits and crops.