Abstract—Similarity feature-based selection and classification (SFSC) algorithms, introduced by Tran et al. in 2013, have been used as a tool to reduce storage cost and increase performance of face recognition systems. However, these still exist a problem when automatically selecting a suitable threshold. This paper introduces a new approach, which combines SFSC algorithms, and a wrapper model, to automatically select a suitable threshold and improve face recognition accuracy. The training face image set (which is split into two separated subsets including a training subset and a wrapper subset) is utilized as data input for the similarity feature-based selection algorithm in combination with the wrapper model to identify a best feature set. The obtained feature set will be used for classification. The experiments were conducted on the histogram-based feature and two databases, ORL database of faces and Georgia Tech face database. The results demonstrated that the proposed algorithm not only allowed for automatic detection of the suitable feature set, but also achieved a better recognition accuracy compared to conventional algorithms. Index Terms—Face recognition, similarity feature, feature selection, filter model, wrapper model. I. INTRODUCTION Feature selection, as an essential task in a face recognition system, could be considered the next step after the feature extraction process [1]. A good dimensionality reduction method can decrease the dimension of feature space, increase recognition accuracy, while maintaining the lowest level of classification errors. A feature selection method selects the best subset of the input feature set that properly describes the given problem with a minimum reduction in performance. Feature selection methods broadly fall into three models: filter, wrapper, and embedded [2]-[4]. The filter model evaluates features without involving any learning algorithm. The wrapper model requires a learning algorithm and uses its performance to evaluate the goodness of features. The embedded model incorporates feature selection as part of the learning process, and uses the objective function of the learning model to guide the process of searching for relevant features such as decision Manuscript received November 20, 2019; revised April 19, 2019. Chi-Kien Tran is with the Faculty of Information Technology, Hanoi University of Industry, No. 298, Cau Dien Street, Bac Tu Liem district, Hanoi, Vietnam (e-mail: chikien.tran@haui.edu.vn). trees or artificial neural networks. In a recent study, two feature selection and classification algorithms based on a filter model that named similarity feature-based selection and classification algorithms (SFSC) have been initially proposed by Tran et al. [5]. The goal of the algorithms is to retain similarity features of the training images in a class in order to minimize within-class differences, while maximizing between-class differences and to use this feature set for classification. They have been proven an efficient tool for improving the performance of face recognition systems using local binary patterns (LBP), local ternary patterns, and local directional pattern (LDP) features [6], [7]. However, SFSC algorithms still have a limitation as the value of threshold parameter is not automatically set, meaning that user needs to test many different values of threshold to find the best similarity feature set. To overcome this limitation, we propose a novel approach based on wrapper model, WSFSC, to find the optimal similarity feature set. Firstly, a face image set is divided into three subsets: training images, wrapper images, and testing images. Secondly, similarity feature-based selection algorithm (SFS) is conducted on two subsets (training images, wrapper images) to find the optimal similarity feature set. Finally, the best similarity feature set is used for classification. The experiments on the ORL database of faces (ORL) [8] and Georgia Tech face database (GTFD) [9], [10] showed that the proposed method was effective for performance improvement of face recognition system. The remaining part of the paper is organized as follows. Section II describes the similarity feature-based selection and classification algorithm. Section III presents SFSC-Wrapper model. The experimental results on the face databases and discussion are presented in Sections IV. Finally, Section V draws the conclusion remarks. II. SIMILARITY FEATURE-BASED SELECTION AND CLASSIFICATION ALGORITHMS The similarity feature-based selection and classification algorithms (SFSC) is an effective algorithm to decrease the dimensions of feature space and improve face recognition rates [5]-[7]. For the similarity feature-based selection algorithm (SFS), its fundamental aim is to retain the similarity features of the training images in a class to minimize within-class differences. First, the variance of features is computed. Next, the obtained values of previous step are normalized. Finally, the features greater than threshold ɛ (threshold value is set by user) are removed. For the similarity Face Recognition Based on similarity Feature-Based Selection and Classification Algorithms and Wrapper Model Chi-Kien Tran International Journal of Machine Learning and Computing, Vol. 9, No. 3, June 2019 357 doi: 10.18178/ijmlc.2019.9.3.810