Hybrid Features based Gender Classification Muhammad Nazir1, Anwar. M. Mirza2 1National University of Computer and Emerging Sciences, FAST AK Brohi Road H-11/4, Islamabad, Pakistan 2Department of Computer Engineering, College Of Computer and Information Sciences, King Saud University, P.O Box. 2454, Riyadh 1154, Saudi Arabia muhammad.nazir@nu.edu.pk, ammirza@ksu.edu.sa Abstract: In this paper, a two step efficient pose based gender classification technique has been presented. In first step, the proposed technique uses different type of features for pose classification which classifies the input image as frontal or side pose image. Based on pose classification result, gender classification is performed in the second step. Two different types of gender classifiers have been proposed for both side pose and frontal images which use the image and pose classification result as input and classifies the image as male or female. Keywords: Pose based Gender classification, Hybrid Feature extraction 1 Introduction A lot of work related to gender classification exists in literature. Chiraz BenAbdelkader et. al. [3] has come up with a technique which uses local region for effective gender classification. Shan et al. 2011 [7] investigated gender recognition on real-life faces using the recently built database, the Labeled Faces in the Wild (LFW). Local Binary Patterns (LBP) was employed to extract features from face, and Ad boost technique was applied to select final set of features from Local Binary Patterns Histogram (LBPH) bins at different scales. For gender classification Support vector machine (SVM) was adopted and boosted LBP features were given to SVM. They obtained the performance of 94.81%. 2 Proposed Method The proposed method consists of two main modules including pose classification and gender classification. Preprocessing module consists of face detection, block based feature selection, feature extraction and classification as in [2]. The proposed technique starts by extracting features using well known feature extraction techniques like Discrete Cosine Transformation (DCT) and Principle Component Analysis (PCA). Features are further processed and classification takes place accordingly. To minimize the computational cost only facial part has been used in the proposed method. This facial part is then used to extract significant features using DCT and Discrete Wavelet - 158 -

Transform (DWT) based reduced features selected through PCA Transform (DWT) based reduced features selected through PCA. It reduces the total number of computations thereby yielding classification in a more robust and efficient manner. Pose classifier is considered as a binary classifier which classifies whether there is a frontal pose in the input image or not based on selected features. After getting the pose information from the pose classifier, gender classification is performed using two different classifiers. One classifier is trained on frontal faces whereas the other one is trained on side poses. The main steps of the proposed methodology are given below: The datasets of images used in the proposed technique as input includes FERET [5], CVL database [1] and Indian database [4]. Images passed to the system are first converted into gray scale and resized to size 100x100. DCT is applied on selected images to extract features and optimal feature selection is performed by selecting the features at top left corner in a zigzag manner [6]. Furthermore, DWT features are also extracted and reduced by PCA have also been used to achieve better accuracy. The new feature vector comprising of most significant features is given to ensemble of heterogeneous classifiers which includes SVM, KNN, LDA, NMC, NMSC and FLDA for the purpose of pose classification. The output of this classification phase determines whether the given input image is a frontal face or face with side poses in either left or right direction. The second module Table 1. Comparison of existing techniques with Proposed method. Technique Dataset KNN LDA NMC Classifier Ensemble Gender Classification using DCT Features CVL Database 0.9746 0.6962 0.9867 FERET 0.9785 0.7745 0.9650 0.9757 Indian 0.9578 0.7865 0.9878 Gender Classification Using DWT Features 0.8350 0.9350 0.8530 0.9830 0.8940 0.9535 0.8830 0.9630 0.9035 0.9450 0.9135 0.9725 utilizes features extracted through DCT and DWT + PCA of both frontal and side pose images and generates the binary output based on majority voting of ensemble based classification system. For the purpose of classification of images of different pose, we need some key parameters of the image to discriminate pose and gender. These parameters representing an image are the feature set of the image. Feature extraction can save computational time and extract useful data that can be used effectively for classification. Input data in the form of features is passed to the classifier for classification. In the first phase the preprocessed input data is fed into the classifier to determine the pose of the subject. It results in either a frontal pose or a side pose. The features of frontal and side pose are fed into the second classification module which determines the gender. Addition of pose classification increases the accuracy of gender classification significantly as against the use of only frontal pose based gender classification. - 159 -

3 Experimental Results 4 Conclusion References In our proposed approach, we have divided input image dataset into two separate subsets used for training and testing. We have used 10-fold cross validation to remove bias. Classifiers used in the proposed technique include BPNN, SVM, NMS, NMC and LDA. Results are given in table shown above. 4 Conclusion In this paper, we have proposed the use of hybrid feature extraction techniques to perform gender classification using ensemble of classifiers. Experimental results show that the performance of the proposed technique is better than the existing state-of-the- art gender classification techniques. References Computer Vision Laboratory, CVL face Database, Retrieved on October 25th, 2011 http://lrv.fri.uni-lj.si/facedb.html. Nazir M., Arfan Jaffar M., Hussain Ayyaz, Mirza, A.M, "Efficient Gender Classification by Optimization of Hybrid Classifiers using Genetic Algorithm", International Journal of Innovative Computing, Information and Control (IJICIC), Volume 7, Number 12, December 2011. BenAbdelkader C., Griffin P., "A Local Region based Approach to Gender Classification", Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05). Indian Institute of Technology, Indian Face Database, Retrieved on October 25th, 2011. http://vis-www.cs.umass.edu/~vidit/IndianFaceDatabase/ National Institute of Standards and Technology, Face Recognition Technology database, Retrieved on October 25th, 2011. http://face.nist.gov/colorferet/ Nazir M., Ishtiaq Muhammad, Batool Anab, Arfan Jaffar Mirza, A.M,, "Feature Selection for Efficient Gender Classification", WSEAS, University of G.Enescu, June 13-15, 2010, Iasi, Romania. Shan, C. Learning local binary patterns for gender classification on real-world face images. Pattern Recognition Lett.(2011), doi:10.1016/j.patrec.2011.05.016. - 160 -