Face recognition using point symmetry distance-based RBF network
Introduction
Face recognition is an important problem with many potential applications, such as surveillance, security, closed circuit television (CCTV) control, film processing and user authentication, etc. Face recognition is a process of identification of a person from a known set of faces stored in a large database. A face identification system having no rejection criteria maps all the input faces presented to the system to the closest known face, giving the identity of the person. However, to make face identification more realistic, some rejection mechanism should also be included into the system. In this case, the system not only recognizes the known faces but also rejects the unknown faces, which are not stored in the database. Therefore, the face recognition system has to accept wide variation of facial expression, head rotation, illumination, background and so on.
A number of different methodologies have been reported since last two decades to tackle the face recognition problem. Samal et al. presented a survey on different approaches used for face recognition [25]. Recently, Yang et al. also presented a survey on face detection approaches [37]. Most of these approaches can be grouped into feature-, template-, principal component analysis (PCA)- and neural network (NN)-based techniques.
Feature-based approaches attempt to extract the invariant salient features from the face, such as eyebrows, eye, nose, mouth, distances and angles between points, intensity distribution and so on. Based on the computed features, a statistical model is built to describe their relationships and to verify the presence of a face. Goudail et al. proposed a method for face recognition using local autocorrelation coefficients [8]. The features were extracted by sliding 25 different kernels over the entire face. Although these features are computationally inexpensive, the performance of the system may drop if the background and lighting condition changes. Phillips used geometrical features for face recognition [22]. Here, the relative positions and shapes of the different features are measured. Many other methodologies based on facial features have also been proposed in the past [2], [11], [31], [38].
The drawback with these feature-based techniques is that the facial features can be severely corrupted due to the changes of illumination and facial expression, and thereby affecting the recognition rate. Lam et al. tried to use pose invariant facial features by using three-dimensional head model [13]. But, they have achieved only 84% correct recognition rate.
In template-based approaches, an input face is compared with a predefined standard face pattern, which can capture the whole face or regions corresponding to the location of facial features, such as face contour, eye, nose, mouth, etc. The recognition of the input face depends on the result of the comparison. Different approaches based on template matching have been proposed [5], [16].
The advantage of this approach is that it is relatively simple to implement. However, the approach is quite ineffective while dealing with faces with variation of scale, pose and shape.
Principal component analysis (PCA) is a statistical method based on the use of Karhunen–Loeve transforms to reduce the dimensionality of a vector by approximating it with most significant eigenvectors, known as Eigenfaces. An image of size n by m can be represented by a vector of size n × m by concatenating the rows of the given image. The idea is that the image vector of size n × m is converted into Eigenvectors, spanning an optimal subspace (face space) such that the mean square error between the projection of the input image onto this subspace and the original image is minimized. In face recognition systems, face images are projected onto the subspace and clustered. In order to recognize a face, the distance between the face image region and the face space is computed. Depending upon the value of this distance, input face image is recognized. Many researchers have used this PCA technique for face recognition. To name a few, Er et al. [6] and Thomaz et al. [33] have used PCA method with radial basis function (RBF) networks for face recognition.
The disadvantage of the method is that the PCA technique retains unwanted variation due to lighting, facial expression and other factors [2].
Neural networks-based approaches are learned from the example-images and rely on the techniques from machine learning to find the relevant characteristics of face images. The learned characteristics, in the form of discriminant functions (i.e. non-linear decision surfaces), are subsequently used for face recognition. Conventionally, face images are projected to a low-dimensional feature space and non-linear decision surface is formed using multilayer neural networks for classifications and recognition [24]. Neural networks have also been used successfully for face recognition problem [6], [19], [24]. The advantage of using the neural networks for face recognition is that the networks can be trained to capture more knowledge about the variation of face patterns, and thereby achieving good generalization [36]. The main drawback of this technique is that the networks have to be extensively tuned to get exceptional performance. Among the neural networks approaches for face recognition, multilayer perceptron (MLP) with back propagation (BP) algorithm has been mostly used [34]. However, the convergence of the MLP networks is slow and the global minima of the error space may not be always achieved [6]. On the other hand, the RBF neural networks have fast learning ability [17] and best approximation property [7]. So, in recent times, many researches have used RBF networks for face recognition [10], [23], [36]. However, their success rates are not so promising as the error rates vary from 5 to 9% under variation of pose, orientation, scale and light [36]. This may be due to the fact that the selection of the centers of the hidden layer neurons might not have been done by capturing the knowledge about the distribution of training patterns and variations of face pose, orientation and lighting.
In the work presented here, an RBF network-based face recognition system has been developed, in which the centers of the hidden layer units are selected by a modified version of the conventional k-means algorithm, which has been recently developed by us [27]. In this algorithm, we have used a heuristic approach to select the initial cluster centers and point symmetry distance instead of Euclidean distance as similarity measure. The Euclidean distance, which is commonly used as similarity measure in the conventional k-means algorithm may fail to detect proper clusters for symmetric objects, like faces, in different poses and orientations. However, the use of point symmetry distance as similarity measure can detect these clusters more properly. Low-resolution face images are applied directly for training as well as testing the system. The performance of our RBF network is promising when applied to the ORL face database.
The rest of the paper is organized as follows. The structures and the learning procedures of the proposed RBF networks are presented in Section 2. The ORL face database and its normalization process have been discussed in Section 3. The experimental results are presented and discussed in Section 4. Finally, in Section 5 conclusion has been drawn.
Section snippets
Architecture of the proposed RBF neural networks
The architecture of an RBF neural network is shown in Fig. 1. It is a feed forward multilayer neural network having one input, one hidden and one output layer. The function of an RBF network can be viewed as follows: firstly, it transforms the non-linearly separable input feature space into a working feature space (usually high-dimensional) by non-linear functions and then fuses the working feature space by linear functions to produce the output space.
The input layer of this network has p
The face database
We have applied our face recognition system with the AT&T Laboratories Cambridge database (formerly called ORL face database) [18]. This database contains 400 grayscale images of 40 persons. Each person has 10 images, each having a resolution of 92 × 112 and 256 gray levels. Images of the individuals have been taken varying light intensity, facial expressions (open/closed eyes, smiling/not smiling) and facial details (glasses/no glasses). All the images were taken against a dark homogeneous
Experiments and results
Many experiments have been carried out, by using several configurations of the proposed system, to determine the accuracy of the face recognition technique. In the following experiments (experiments 4.1.1–4.1.5), five images from each individual in the database are selected randomly for training set and the rest of the face images are included in the test set. Therefore, a total of 200 faces are used to train and another 200 faces are used to test the RBF networks. It should be noted that there
Conclusion
A face recognition system using an RBF neural network, whose hidden layer units are modeled by using a heuristic approach and point symmetry distance as similarity measure, has been described. The system has been tested on the AT&T (formerly ORL) face database after sub sampling each image with a resolution of 16 × 16 and 256 gray levels [26], [36]. The performance of the present method has been evaluated with two experimental methodologies; one with no rejection criteria and another having
Acknowledgements
This work was partially supported by the Center for Microprocessor Applications Training Education and Research (CMATER) and Storage Retrieval and Understanding of Video for Multimedia (SRUVM) projects of the Dept. of Computer Science & Engineering, Jadavpur University, Kolkata 700 032, India.
References (39)
- et al.
Face recognition approach based on rank correlation of gabor-filtered images
Pattern Recog.
(2002) - et al.
Learning identity with radial basis function networks
Neurocomputing
(1998) - et al.
Human facial feature extraction for face interpretation and recognition
Pattern Recog.
(1992) - et al.
Face recognition based on the uncorrelated discriminant transformation
Pattern Recog.
(2001) - et al.
A hierarchical multiscale and multiangle system for human face detection in a complex background using gravity-center template
Pattern Recog.
(1999) - et al.
Face recognition using transform features and neural networks
Pattern Recog.
(1997) - et al.
Automatic recognition and analysis of human faces and facial expressions: a survey
Pattern Recog.
(1992) - et al.
Connectionist models for face processing: a survey
Pattern Recog.
(1994) - et al.
Feature-based human face detection
Image Vis. Comput.
(1997) - et al.
Eigenfaces versus fisherfaces: recognition using class specific linear projection
IEEE Trans. Pattern Anal. Mach. Intell.
(1997)
Neural Networks for Pattern Recognition
Face classification using a multiresolution principal component analysis
Face recognition: features versus template
IEEE Trans. Pattern Anal. Mach. Intell.
Face recognition with radial basis function (RBF) neural networks
IEEE Trans. Neural Netw.
Networks and the best approximation property
Biol. Cybern.
Face recognition system using local autocorrelations and multiscale integration
IEEE Trans. Pattern Anal. Mach. Intell.
Neural Networks a Comprehensive Foundation
An analytic-to-holistic approach for face recognition based on a single frontal view
IEEE Trans. Pattern Anal. Mach. Intell.
Face recognition: a convolutional neural-network approach
IEEE Trans. Neural Netw.
Cited by (46)
Interval type-2 fuzzy set induced fuzzy rank-level fusion for face recognition
2023, Applied Soft ComputingShort-term load forecasting using SVR (support vector regression)-based radial basis function neural network with dual extended Kalman filter
2013, EnergyCitation Excerpt :An annealing clustering ANN [17], a BP (back propagation) neural network with rough set [18], a neural network combining wavelet transform and adaptive mutation particle swarm optimization [19], an ANN based on fuzzy logic methods [20], and Bayesian neural networks integrating Monte Carlo algorithm [21] have been proposed for short-term load forecasting. Among ANNs, the radial basis function neural network (RBFNN) has received considerable applications in many fields [22–25]. Since RBFNN has only one hidden layer and has fast convergence speed.
Integration of support vector regression and annealing dynamical learning algorithm for MIMO system identification
2011, Expert Systems with ApplicationsCitation Excerpt :A neural inverse dynamic NARX model has been adopted to perform MIMO system identification (Anh & Phuc, 2010). In the neural network, RBFNs have received considerable applications, such as function approximation, prediction, recognition, etc. (Chuang, Jeng, & Lin, 2004; Sing, Basu, Nasipuri, & Kundu, 2007; Xu, Xie, Tang, & Ho, 2003; Yu, Gomm, & Williams, 2000). Since RBFNs have only one hidden layer and have fast convergence speed, they are widely used for nonlinear system identification recently (Apostolikas & Tzafestas, 2003; Chen, Hong, Luk, & Harris, 2009; Falcao, Langlois, & Wichert, 2006; Fu et al., 2009; Li & Zhao, 2006).
Face recognition by generalized two-dimensional FLD method and multi-class support vector machines
2011, Applied Soft Computing JournalCitation Excerpt :To utilize higher order statistics, some nonlinear forms of eigenface and fisherface methods have been developed [27–32] for better recognition performance. The advantage of using the neural networks for face recognition [11,15–19] is that the networks can be trained to capture more knowledge about the variation of face images and thereby achieving good generalization. In recent times, among the neural network approaches, many researchers have used RBF neural networks (RBFNN) for face recognition [11,15–19].
Radial basis function networks with hybrid learning for system identification with outliers
2011, Applied Soft Computing Journal