I.J. Image, Graphics and Signal Processing, 2017, 5, 68-75 Published Online May 2017 in MECS (http://www.mecs-press.org/) DOI: 10.5815/ijigsp.2017.05.07 Copyright © 2017 MECS I.J. Image, Graphics and Signal Processing, 2017, 5, 68-75 Pose Normalization based on Kernel ELM Regression for Face Recognition Tripti Goel, Vijay Nehra Department of Electronics and Communication Engineering, Bhagat Phool Singh Mahila Vishwavidyalaya, Khanpur Kalan, Sonepat, Haryana, India Email: triptigoel83@gmail.com, nehra_vijay@yahoo.com Virendra P. Vishwakarma University School of Information and Communication Technology, Guru Gobind Singh Indarprastha University, Dwarka, Delhi, India Email: virendravishwa@rediffmail.com Abstract—Pose variation is the one of the main difficulty faced by present automatic face recognition system. Due to the pose variations, feature vectors of the same person may vary more than inter person identity. This paper aims to generate virtual frontal view from its corresponding non frontal face image. The approach presented in this paper is based on the assumption of existence of an approximate mapping between the non frontal posed image and its corresponding frontal view. By calculating the mapping between frontal and posed image, the problem of estimating the frontal view will become the regression problem. In the present approach, non linear mapping, kernel extreme learning machine (KELM) regression is used to generate virtual frontal face image from its non frontal counterpart. Kernel ELM regression is used to compensate for the non linear shape of the face. The studies are performed on GTAV database with 5 posed images and compared with linear regression approach. Index Terms—Face recognition, Kernel Extreme Learning Machine Regression, Pose normalization, Virtual Frontal View. I. INTRODUCTION Face recognition is used extensively for biometric identification since last two decades. The popularity of face recognition system is due to its advantages of being passive and non-intrusive nature. It also provides higher recognition accuracy as compared to other biometric identification techniques. However, it suffers from serious challenges under outdoor environments, for example, appearance of the face may vary too much due to different poses. In current reviews [1,2], pose variation is identified as one of the main unsolved problem for face recognition system. Therefore, it has attracted the interest of many researchers to normalize the pose variations. A lot of approaches have been proposed to deal with recognizing faces under different poses. View based methods [3-8] were mainly used for pose normalization, but it usually requires multiple face images of each subject with different poses. 3D model based methods [9- 20] are also explored to normalize the pose variations, but these methods are too slow to use them in real time scenario. Generating virtual frontal view from its non frontal view [21-28] is one of the popular solution to normalize pose for face recognition. By generating the virtual frontal view of the posed image, either, all face images are normalized to the frontal view or gallery can be extended to cover the large pose variations. Local linear regression (LLR) method is proposed by Chai et al. [24] for efficiently generating the virtual frontal view for the posed face image. In this method, the posed image is partitioned into multiple patches and then linear regression is applied to each patch for predicting its corresponding virtual frontal patch. LLR method to normalize pose variations is simpler as well as easier for real world tasks as only linear regression has to be done. Also, it requires only the coarse alignment based on the center of the two eyes. However, main drawback of linear assumption is the loss of lots of information, as the rotation of the human head is a non linear problem. Therefore, to deal with non linearity of rotation of human head, there is the need of non linear regression to solve the pose variations. In this paper, kernel extreme learning machine (KELM) regression [31-34] is proposed to efficiently estimate the virtual frontal face image from its non frontal view. Kernel ELM is used here as the face image is non linear in shape and the different poses change non linearly. Linear regression cannot estimate these non linear changes accurately. Compared to linear regression, KELM regression is more efficient for virtual view generation as it considers non linear shape of the face view. The remaining paper is organized as: Section 2 gives the review of existing techniques for pose normalization. Section 3 explains the linear regression for virtual view generation. Section 4 explains the KELM regression for generating the virtual frontal face view. Section 5 presents the results on GTAV face database. The conclusion and further recommendations for pose