Spectral Eigenface Representation for Human Identification

Human identification based on face images, as physical biometric means, plays an imperative role in many applications area. The methods for human identif ication using face image uses either part of th e face, all face, or mixture from these methods, in either time domain or frequency domain. This paper investigate the ability to implement the eigenface in frequency domain, the res ult spectral eigenface is utilize as a feature vector means for human identification. The co nverting from eigenface implement ation in time domain, into spectral eigenface implementation in frequency domain, is based on implemented the correlation by using FFT. The Min-max is invoked as normalizat ion techniques that increase s pectral eigenface robustness to variations in facial ge ometry and illumination. Tw o face images are contrast in t erms of their correlation distance. A threshold (10.50x107) is used to restrict the impostor face image from being identified. The experimental results point up the effectiveness of a new method in either using varying (noisy images , unknown image, face expressions, illum ine, and scales ), with identification value of 100%.


Introduction
Biometrics is fundamentally concerned with digitally encoding physical characteristics of the user's voice, eye, hand or face to a single ID.Biometrics applications require reliable and automatic personal identification for effective application requirements.Traditional, automatic, personal identification can be divided into two categories: token-based, such as a physical key, an ID card and a passport, and knowledge-based, such as a password.However, these approaches have some limitations.In the token-based approach, the "token" can be easily stolen or lost.In the knowledge-based approach, to some extent, the "knowledge" can be guessed or forgotten.Thus, biometric personal identification is emerging as a powerful means for automatically recognizing a person's identity [1].Identification can be applied in a closed system such as employee positive identification for building access, or in an open system such as a national ID system.Positive biometric identification, a 1-to-many problem, is more challenging than verification, a 1-to-1 problem.As stated in, "positive identification is perhaps the most ambitious use of biometrics technology " 1) is a simple biometric system has four important components and two phases [1]: 1. Sensor: This captures the biometric data of an individual.2. Feature Extraction: is the stage in which the acquired data is processed to extract feature values.3. Matching phase: is carried out when the feature values are compared against those in the template by generating a matching score.4. Decision-making phase is done when the user's claimed identity is either accepted or rejected based on the matching score generated in the matching module.The enrolment phase: is responsible for enrolling an individual into the biometric system.During the enrolment phase, biometric characteristics of an individual are first scanned to produce a raw digital representation of the characteristics.In order to facilitate matching, a feature extract to generate a compact but expressive representation, called a template, further processes the raw digital representation of the template may then be stored in a central database or recorded on a magnetic card or smart card (and issued to an individual) depending on the nature of the application.The identification phase: is responsible for identifying individuals at the point of access.During the operation phase, the biometric reader captures the characteristics of individuals to be identified and converts it to a digital format, which is further processed by the extractor to produce the same representation.The ensuing image is fed to the feature matcher who compares it against the templates to establish the identity [1].
PDF created with pdfFactory Pro trial version www.pdffactory.com The following factors are needed to have a successful biometric identification method [1]: 1.The physical characteristic should not change over the course of the person's lifetime 2. The physical characteristic must identify the individual person uniquely 3. The physical characteristic needs to be easily scanned or read in the field, preferably with inexpensive equipment, with an immediate result 4. The data must be easily checked against the actual person in a simple, automated way.Other characteristics that may be helpful in creating a biometric identification scheme are[1]: 1. Ease of use by individuals and system operators, 2. The willing (or knowing) participation of the subject is not required Face identifications fill into three categories: holistic methods, which use the whole face image for recognition, featurebased methods, which use local regions such as eyes or mouth, and hybrid methods, which use both local regions and the whole face.In spite of this, face identification technology seems to be a difficult task to develop since the appearance of a face varies dramatically because of illumination.Facial expression, head pose, and image quality determine the recognition rate.In addition, the number of the same face in the database with different facial expression should be sufficient so that the person can be identified in all possible situations [3].Many reserch on face identification based on Principal Component Analysis (PCA).However, any system in this world has its limitations and can be improved.To overcome the disadvantages of PCA, such as large computational load and low discriminatory power it can be combined with other techniques.The objective of this paper is to present a novel spectral eigenface implementation, as proposed means for human identification.A threshold is used to restrict the impostor face image from being identified.The experimental results point up the effectiveness of a new method PDF created with pdfFactory Pro trial version www.pdffactory.com in either using different) noisy images, unknown image, face expressions, illumine, and scales ) .
In the following subsections, background knowledge is presented in section 2. Secondly, the purposed spectral eigenface is in Sections 3. In Section 4, Excrement results, followed by a conclusions in Section 5.

Background 2.1 Discrete Time Fourier Transform
Once the signal has been acquired and digitized, it can be converted to the frequency domain by using Fast-Fourier-Transformed (FFT).The FFT results can be either real and imaginary, or magnitude and phase, functions of frequency.The choice of output format belongs to the user.the transform and inverse transform pair given for vectors of length N by [4]: is the Nth root of unity.

Correlation Implementation Using FFT
Let X, and Y be data sets such that, correlations based FFT is defend as: take FFT of X, and FFT of Y, multiply one resulting transform by the complex conjugate of the other, and inverse transform the result product such as [5]: 6. Compute the eigenvectors ui of AAT : PDF created with pdfFactory Pro trial version www.pdffactory.com6.1 Consider matrix AAT as an MxM matrix.6.2 Compute the eigenvectors vi of AAT such that where 6.3 Compute the M best eigenvectors of AAT:

Normalization
A feature is normalized by scaling its values so that they fall within a smallspecified range, such as 0.0 to 1.0.For distanced-based methods, normalization helps prevent features with initially large ranges from outweighing features with initially smaller ranges.
Min-max normalization performs a linear transformation on the original data.Suppose that mina and max are the minimum and the maximum values for feature A. Min-max normalization maps a value v of A to v′ in the range [newmin, newmax] by computing:

Eigenface Matching
Let X and Y be two spectral eigenface feature vectors where, xi ∈ X, yi ∈ Y, i=1,…,n.to calculate the degree of association, a correlation distance is defined as[5]: where r is the linear correlation coefficient which is given by the formula [5]: where _ x is the mean of the vector X, _ y is the mean of the vector Y. the correlation distance determines the genuine or forged query sample, it is easy to verify the input pattern by a pre-defined threshold value T. If the value R is smaller than threshold T, then the owner of query sample is claimed to be individual X.Otherwise, the query sample is classified as a forged pattern.

Threshold Selection
In any face identification system, it is essential to pike a suitable threshold (T), for a good performers results.To this end, an approach based on intra-class and interclass information collected from the Enrolment database.The intra-class (D) measures the distances between images of the same individual, therefore it gives an indication of how similar the images of the same individual are.The intra-distance is defined as: , and PDF created with pdfFactory Pro trial version www.pdffactory.comStep6: End.

Eigenface Identification Process
An unknown query face image can be represented as a linear combination of the best K Eigenfaces of the obtaining eigenvectors for a given dataset.In face identification, the eigenfaces are used once again in order to compute a distance from the query image in the face space, as presented in

The proposed Spectral Based Eigenface Feature
The new proposed idea of applying the FFT in the implementation for Eigenface is by use FFT in the implementation of the correlation, as an alternative of conventional ideas of converting the intensity of the image face data into the spectral domain, followed by applying the Eigenface.The new proposed idea is named as the spectral based eigenface feature.

PDF created with pdfFactory Pro trial version www.pdffactory.com
The correlations can be computed by using the FFT as follows: FFT the two data sets, multiply one resulting transform by the complex conjugate of the other, and inverse transform the product.Hence, by evaluating this cross correlation, a speed up ratio can be obtained comparable to conventional Eigenface.Let I denote a

Experiment Results
A set of faces full between April 1992 and April 1994 at the Olivetti Research Laboratory (ORL) in Cambridge, UK.There is 10 different images of 40 distinct individuals.The images were taken at different times, varying illuminance, facial expressions "open/closed eyes, smiling/non-smiling " and facial details "glasses/no-glasses,"", a sample of the used images is in Figure (3).All the images are taken against a dark homogeneous background and the individuals are in up-right, frontal position "with acceptance for some side movement ".The files are in PGM format, with a size of each image is 92x112, 8-bit grey levels.The Enrolment Database consists of 160 normalized spectral eigenfaces, 4 for each individual.A threshold T is calculated for the enrolment database, to prevent unknown individual from being identified and found to be equal to 7 10 50 .10 × .The rest 360 normalized spectral eigenfaces, 6 for each individual, is used in the identification phase, the results are depicted in Table (2).The result of any identification goes in one of four situations.A correct identification: identify an individual already registered in the enrollment database.A correct refusal: refuse an individual not registered in the enrolment database.A wrong acceptance: accept an imposter not registered in the enrolment database, or to identify imposter as someone in the database incorrectly.Wrong refusal: refuse a genuine user registered in the enrolment PDF created with pdfFactory Pro trial version www.pdffactory.comdatabase, or to identify the genuine users as unknown incorrectly.The performance of the spectral eigenfaces approach is study under different conditions either from the 360 rest images or out of it, as depicted in bellow subsections.Identification With Different Head Tilts: The robustness of the spectral eigenfaces identification algorithm to head tilt is studied, with different head tilts either leftoriented or right-oriented, top-orientation, and down-orientation as shown in Figure (4(a-e  The robustness of the spectral eigenfaces identification algorithm is studied, with different unknown face images for boys and girls with glass and without, with varying faces expressions, as shown in . The result of the identification is depicted in Table (5).

Conclusions
this paper investigate the ability of implementing the eigenfces in the frequency domain by using the FFT as a means for Human identification.The ORL face images are used to evaluate the performance of new proposed algorithm, which is implemented by using Matlab 7 as programming language.The proposed eigenface based FFT system is investigated and it found that: 1. the used of min-max approch as a normalization method for the result feature vector is to remove the outliner in the enrolment database.2. The benefit of using a threshold is to prevents the impostor from being identified.3.One of the major advantages of spectral eigenfaces recognation approch is the ease of implementation.Futhermore, no knowledage of geometry or specific feature of the face is required, and only a small amount of work is needed regarding preprocessing for any type of face images.PDF created with pdfFactory Pro trial version www.pdffactory.com

PDF created with pdfFactory
The inter-class (P): The distances between the images of an individual are measured against the images of other individuals in the Enrolment database, therefore it gives an indicates how different each image of an individual is when compared to images of other individuals in the Enrolment database.The inter-distance is defined as: T) is then calculated from intra-class and inter-class information as described in[2].The estimation of threshold (T) depends mainly on the number of images per individual in Enrolment, therefore as in[3], every individual should have at least 4 images for Enrolment database.The algorithm for the maximum intra-class and the minimum inter-class calculation is described as: the intra-class (D), and the inter-class (P) in ascending order Step5: Compute Dmax, and Pmin .

Figure
then x is a face.De is the distance from face space.
be seen as a point in Rn .Here are the steps to computing these Eigenfaces: 1. Obtain face images I1 , I2 ,..., I M (training faces).2. Represent every image Ii as a vector xi .the eigenvectors vi of AAT such that: )). Identification with Varying Illuminance: The robustness of the spectral eigenfaces identification algorithm to head tilt is studied, as depicted in Figure(5(a-c)) with face images moved by 45 degrees and the other with light moved by 90 degrees.Identification with Varying Head Scale: The robustness of the spectral eigenfaces identification algorithm studied, with a medium head scale and the other with a small one, as shown in Figure(6(a-c)).Identification With Different Face Expression: The robustness of the spectral eigenfaces identification algorithm is studied, with different head face expression either smiling, eye move, with glasses, and smiling with glasses as shown in Figure(7(a-e)).Identification with Different Noise Type and Level: The robustness of the spectral eigenfaces identification algorithm over noise image, with different noise types as shown in).

Figure
Figure (3): A sample of the ORL Face Images

Table (
Pro trial version www.pdffactory.com