Individual Recognition System using Deep network based on Face Regions

International Journal of Applied Mathematics Electronics and Computers ISSN:2147-82282147-6799 http://dergipark.gov.tr/ijamec Original Research Paper Individual Recognition System using Deep network based on Face Regions Abdelouahab ATTIA*1, Mourad CHAA2 Accepted : 01/08/2018 Published: 30/09/2018 DOI: 10.1039/b000000x Abstract: Biometric based face recognition is a successful method for automatically identifying a person using her face, with a high confidence. For that reason, this paper introduces an efficient method for face recognition based on deep networks. It considers the three face regions: eye, mouth, and face. First, we have built one sparse autoencoder for every single region their outputs will be concatenated together and fed into another sparse autoencoder. After that, the softmax layer has been employed in the classification step. However, with a deep network method known as the softmax layer has been formed by stacking the encoders from the autoencoder. Followed by formed the full deep network. Finally, the results have been generated on the test set based on the deep network. In the experimental stage, the Yale B database and the AR database and JAFFE database have been used to test the proposed individual recognition system. Experimental findings have clearly proven that the performance of the introduced algorithm is very encouraging and can respond to the security requirements. Keywords:Deep network; sparse autoencoder; hybrid face regions; individual recognition system; face recognition. 1. Introduction During the past few decades, face recognition it has become an important research field including computer vision, machine learning as well as pattern recognition. Several activities in this field rely on its applications including different areas such as security, criminal identification, surveillance, commercial and credit card verification. However, Facial recognition presents numerous advantages more than other biometric technologies: it is natural, non-intrusive and easy to use. The performance of a person face Recognition system design depends on the step of features extraction that is an essential operation before applying an algorithm of classification. Various methods have been developed to extract facial features. Mainly, the existing methods in the literature are classified into two categories: in the holistic methods including the Principal Component Analysis (PCA) [1], Linear Discriminant Analysis (LDA) [2], Laplacian faces [3], Multilinear Principal Component Analysis (MPCA) [4], and Independent Component Analysis (ICA) [5]. In the appearance based methods, one of them that have been generally used to extract information appearance that based on the Gabor wavelet [6, 7]. However, the major limitation of Gabor filters is a very large number of features that it generated, it is time-consuming as well it requires convolving face images with a Gabor filter bank in order to extract multiscale coefficients and multi-orientation. Recently, methods of texture description and classification are known as Local Binary Pattern (LBP) [8] and local quantization phase (PLQ) [9] have been studied to extract facial appearance-based features. a new _______________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ 1 Computer Science Department - Faculty of Mathematics and informatics Mohamed El Bachir El Ibrahimi University of Bordj Bou Arreridj, 34000, Algeria 2 Lab. ELEC – Faculty of new technology of information and communication Ouargla university,Ouargla 30 000, Algeria * Corresponding Author:Email: decorrelated neural network ensemble algorithm for face recognition has been introduced in [10]. However, in this technique, the 2D-FNNs have been used as base components and the negative correlation learning (NCL) scheme is united with the two-dimensional (2D) feed-forward neural networks (2DNNRW) algorithm to construct ensembles (DNNE 2D-NNRW). The current paper introduces a novel technique for face recognition based on the deep network. Before features extraction techniques step, first the detection of significant information in the face image which captured in other word select the face itself. However, hair, background, and skin have not influence in steps of analysis and identification. Then, the two regions from the face: eye and mouth are extracted from the face selected image. After that, sparse autoencoder (SAE) has been constructed for each distinct region (face, eye, and mouth). The outputs of SAE have been combined together and fed into another sparse autoencoder. Then, the softmax layer has been employed to classify the obtained feature vectors. It is worth to mention that a deep network is formed by stacked the encoders from the autoencoder with the softmax layer. Also with the full deep network formed, the results on the test set have been computed. Finally, the individual recognition system has been tested on the three useful databases Yale B, AR, and JAFFE. The rest of this paper is organized as follows: section two, describe the feature extraction method. And the classification process used in the proposed system, including a brief description of the Softmax classifier and Fine-tuning. Experimental results are given and discussed in section three. The conclusion and further works are drawn in section four. 2. Proposed method 2.1. Feature Extraction The aim of the current work is the use of the autoencoder so that to extract features vectors from a different region of face cited International Journal of Applied Mathematics Electronics and Computers (IJAMEC) 2018, 6(3), 27–32 |27 above. However, an autoencoder is an unsupervised learning algorithm refers to an artificial neural network that is trained to reconstruct its input for its output (encoding). Recently, an autoencoder has been used as a tool for dimensionality reduction, image representation and learn deep neural network [11]. Architecturally, the autoencoder is made up of two main parts: the encoder and the decoder. In the current study, three layers have been used: an input layer, an output layer, and one hidden layer. Given the input to an autoencoder be an image x ∈ ℝD , and the encoder maps of the images x to a different data z ∈ ℝQ as given in the following formulae: 𝑧 (1) = 𝑓 (1) (𝑤 (1) 𝑥 + 𝑏 (1) ) (1) bi are the i row of the weight matrix and the i entry of the bias vector respectively. The parameters of the sparse autoencoder have been empirically selected as α = 0 ∙ 5,β = 10. To form the feature vector based on the proposed method, features have been extracted from face and eyes and concatenated with mouth features. As clearly shown in Figure.1, the obtained feature has been fed into another sparse auto (...truncated)