Feature Selection for the Stored-grain Insects Based on PSO and SVM

Abstract

The feature subset selection is a key preprocessing part in the detection of the stored-grain insects based on the image recognition technology. According to the global optimization ability of the particle swarm optimization (PSO) and the superior classification performance of the support vector machines (SVM), this study proposed a method based on PSO and SVM to improve the classification accuracy with the appropriate feature subset. The single objective fitness function was designed to evaluate the feature subset by introducing the v-fold cross-validation training model accuracy and the number of the selected features. Nine species of the stored-grain insects spoiled seriously in grain-depot, like Tenebroides mauritanicus(L.) and Rhizopertha dominica Fabricius. The feature subset selection for the stored-grain insects was implemented by the method based on PSO and SVM. The optimal feature subset consisted of seven features was selected from the 17 morphological features, such as area and perimeter. Compared with the genetic algorithm (GA), the method in this study can decrease the size of the feature subset and improve the classification accuracy. Making use of the feature subset selected by PSO and SVM, the ninety image samples of the stored-grain insects were classified by the SVM classifier that two parameters had been optimized, and the classification accuracy was over 95.5%. The experiment showed that it was practical and feasible.