Enhanced Image Classification by Eliminating Outliers with the Combination of Feature Selection and K-means Techniques

Abstract

Accurate image classification will yield valuable information to support decision-making. Support Vector Machine (SVM) is a widely used technique to achieve high classification accuracy. However, data outliers can reduce the SVM’s accuracy. To resolve this problem, the K-Means clustering method is used to eliminate the outliers by checking the proximity between data and clustering the data. Nevertheless, one of the challenges of using K-Means is the sensitivity of the initial centroid selection which is done randomly. Therefore, this study combines the use of K-Means, feature extraction with VGG-16 deep learning architecture, and feature selection using the Chi² technique to get better classification accuracy. The combination of these methods is empirically proven to increase the accuracy of three image dataset about 20%. The results demonstrate that using these methods in conjunction can also reduce the amount of time needed for image classification. Nevertheless, label information is not taken into consideration in this study. Therefore, in the future, this research can still be developed by applying other standards and adding information labels in the feature selection process.