Variance-Based Geometric Feature Selection for Face Recognition System

In a feature-based facial recognition system, geometric features are usually combined with other features to produce reliable facial recognition system. There are thousands of geometric features on a face. The difficult part in using geometric feature is determining how many features and choosing which features to use. The best feature selection method is to tests all feature combinations in the system, which is called Holistic Testing, but it is impossible to run this method because of the huge number of feature combination. Several papers use geometric features without performing feature selection process, resulting a less optimal facial recognition system. This paper proposes a feature selection method using variance value approach. The proposed method is applied to 14-point facial landmark dataset based on the FRGC 3D facial database. The variance value is used to select candidate features that make the holistic testing feasible to performed. Based on our experiment, the highest recognition accuracy is obtained in the use of 27 features, they give 92% of recognition accuracy. The whole features selection requires around 8768 second to complete. This paper also compares the selected features used in some previous papers with our selected features. By applying the features in our recognition system, our features give better performance with the same dataset. At the end, this paper concludes that, in order to obtain a good accuracy result, it is important to perform a feature selection process before using them in facial recognition system.