A Robust High Performance Face Recognition using Decorrelation of Local Features by Discrete Cosine Transforms

This paper proposes a novel method of face recognition using de-correlation of local features using Discrete Cosine Transforms (DCT). The impulse for the proposed idea is with the fact that histograms DC constituent of local Gabor binary patterns constitute low frequency components which will sparsely help in actual recognition, because information in face resides in low frequency bands and is similar to all the images and also when these histograms are concatenated, it becomes difficult to differentiate and segregate actual frequency variations which add value for accurate recognition. A high correlation exists in between these histograms. This high correlation affects the recognition accuracy, hence de-correlation is achieved with the help of DCT application for individual histogram bins which aids in identification of actual frequency variations and highlights the changes in between two histograms thus improving the recognition accuracy. This method employs a non-statistical procedure which avoids training step for face samples thereby avoiding generalizability problem which is caused due to statistical learning procedure. The performance modeling is carried out by varying both internal and external factors of face recognition system and improvement is shown considerably high in terms of recognition accuracy and reduction in storage space by storing train images in compressed domain.


INTRODUCTION
In image analysis, face recognition is one of the most representative applications. In law enforcement and security applications, it has gained major importance. Face as a biometric is most popular because face image of a particular subject can be taken in a non-intrusive way without their co-operation. In general, problem statement of automated recognition of faces can be formulated as "Given still or video image of a scene, identify or verify one or more persons in the scene using a stored database of faces" [17]. Solution to the above problem statement involves segmentation of faces -face detection from cluttered scenes as a first step followed with feature extraction from the face regions and recognition/ verification depending upon the application it is used for. Face recognition flow chart is depicted in Fig. 1.

Fig. 1 A generic Face recognition system
In face recognition, crucial step is the assessment and benchmarking of algorithms. Although many face recognition techniques have been proposed and have shown appreciably promising results, design of robust face recognition system is still an open challenge. Illumination variations, pose/ alignment and recognition in outdoor imagery are three major challenges which affects the performance of face recognition. For example, pose discrimination is not that complex compared to accurate pose estimation. In addition to these two problems, there are bigger challenges such as, recognition of a person from images acquired years apart, involving variations in the overall appearance. The performance of face recognition system is limited due to various challenges such as modeling and normalization of illumination, pose variance / pose estimation, accurate landmark detection, feature extraction and model building, matching methodology / determination of distance measures, vast database to be handled (in most applications), subspace learning and so on. The performance of face recognition systems in different stages is depicted in Fig. 2. The factors which are affecting face detection phase are pose, alignment of face, number of frames in video input, background and so on. In feature extraction phase many factors like shape of the eyes, eyebrows, texture of skin, shape of the face needs to be considered strictly. In actual recognition, many challenges arise like rate of recognition, storage space for datasets and complexity of algorithm and so on. This paper proposes non-statistical method for face recognition which not only is robust to the variations of imaging condition but also embeds much discriminating power. It uses effective Gabor filtering of local features and forming histograms as the primary step. Successively DCT operator is applied for feature vector and thereby recognizing faces using similarity measures. It combines local intensity distribution and considers maximum values, which aids in segregating high frequency variations, making this approach a robust recognition technique. Also, storing train images in compressed domain solves security concerns of image splicing and aids in storage space reduction for the datasets. The performance of the proposed method is compared to the state-of-art method and results are documented in experimentation summary section.
The rest of the paper is organized as follows. Section II outlines literature survey, Section III describes the application of Gabor filters to face images and formation of local histogram feature vectors. Section IV details about Local Binary Patterns for GMPs, Section V describes about image compression using Discrete Cosine Transform (DCT) technique, Section VI provides the details of the proposed method of face recognition using DCT coefficient vectors. Experimental results and performance modeling based on the analysis of accuracy of recognition are discussed in Section VII, ending with conclusions in Section VIII and references.

LITERATURE SURVEY
In the course of studying papers which are referenced [2] through [7] in existing state-of-art literature, the concepts such as face image variations caused by different poses, inter-person differences used in distinguishing identities, related potentials in many applications dealing with uncooperative subjects, full power of face recognition as a passive biometric technique, problems due to varying illumination, generalizability problems, problems related to image splicing, performance limitations caused due to vast datasets have been understood.
Previous work carries many face representation approaches, most of them are based on statistical learning, such as linear discriminant analysis, Support Vector Machines or AdaBoost technique which inherently suffers from the generalizability problem due to the difference in between unpredictable distribution of the "testing" face images from real world and the "training" face images. In subspace approaches, a training set is developed in order to extract the most discriminating representation for final classification, but the discriminant subspace is greatly dependent on the training set [17]. If in case the testing images are captured under different environment from those of the training ones, already developed discriminant subspace would be inapplicable to those probe faces.
Though, generalizability problem has been theoretically discussed through Structural Risk Minimization (SRM) principle [1], in most real-world face recognition applications, generalizability problem remains as a huge bottleneck for most face recognition systems. Though there are many methods available for face recognition which avoids generalizability problem such as histogram method, multi resolution histogram method, Gabor feature based method, all these methods has its own pros and cons as listed in Table.1. A thought process map for generalizability problem is depicted in Fig.3

Fig. 3 Feature descriptors which avoids generalizability problem
Currently, many methods in the existing literature focus mainly on actual recognition algorithms. But, face images are extremely sensitive to illumination variation. This paper shows that, if the same importance is given to image preprocessing steps such as cropping and illumination normalization, a better performance can be achieved for any given face recognition system. Also, in certain applications such as airport surveillance systems, the systemic issues and computational complexity involved in storing huge train datasets poses a major challenge in performance of real-time face N o v 1 5 , 2 0 1 3 recognition systems. Although there are commercial systems available in the market, it is still an active and challenging topic, due to the fact that the appearance of the same face looks dramatically different in uncontrolled environments with rich variations of pose, expression, illumination and occlusion.

GABOR MAGNITUDE PICTURES FOR FACE IMAGE
This section describes the filtering of face image and formation of GMPs using Gabor filters. Starting with the introduction of Gabor filters and its construction, the basic concepts of feature extraction for a face using the Gabor filter bank is explained and proceeds with the presentation of the Gabor Magnitude Picture (GMP of a face) representation. Gabor filters which are also called as Gabor wavelets or kernels have been proven to be efficient tool for facial feature extraction and robust face detection and recognition. They represent complex band-limited filters with an optimal localization in both the spatial and frequency domain. Hence, when employed for facial feature extraction, they extract multiresolutional, spatially local features of a confined frequency band [2]. Like all filters operating in the scale-space, Gabor filters also relate to the simple cells of the mammalian visual cortex and hence relevant from the biological point of few as well. In general, the family of 2D Gabor filters can be defined in the spatial domain as follows [8]: Each Gabor filer represents a Gaussian kernel function modulated by a complex plane wave whose center frequency is fp and orientation is θq. The parameters "c" and "η" determine the ratio between the center frequency and the size of the Gaussian envelope. While different choices of the parameters determining the shape and characteristics of the filters define different families of Gabor filters, the most common parameters used for face recognition are c = η =√2 and N o v 1 5 , 2 0 1 3 fmax=0.25 [9,10,11]. When using the Gabor filters for facial feature extraction, researchers typically construct a filter bank featuring filters of five scales and eight orientations, that is, p = 0, 1, . . . , r − 1 and q = 0, 1, . . . , s − 1, where r = 5 and s = 8. The real parts of the entire filter bank will be used for facial feature extraction (comprised of 40 filters) as shown in Fig.4.

Fig. 4 Real parts of the Gabor filter bank.
Let I(a, b) stand for a grey-scale face image of size 128x128 pixels and gp,q(a, b) denote a Gabor filter given by its center frequency is fp and orientation is θq. The feature extraction is defined as a filtering operation of the given face image I(a, b) the Gabor filter gp,q(a, b) as shown below.
Gp,q(a, b) is decomposed into real and imaginary parts as Re[Gp,q(a, b)] and Im[Gp,q(a, b)] respectively. Based on this, magnitude and phase responses are given by following equations.

LOCAL BINARY PATTERN FOR GABOR MAGNITUDE PICTURES OF FACE IMAGE
Local Binary Pattern (LBP) was firstly designed for texture classification. Ahonen et al. [12] successfully applied it to represent faces. LBP considers both shape and texture information to represent the face images. In this approach the textures of the facial regions are locally encoded by the LBP patterns while the whole shape of the face is recovered by the construction of the face feature histogram. The idea behind using the LBP features is that the face images can be seen as composition of micro-patterns such as flat areas, spots, lines and edges which are invariant with respect to monotonic grey scale transformations [13,14]. Combining these micro-patterns, a global description of the face image is N o v 1 5 , 2 0 1 3 obtained. The basic LBP operator labels the pixels of an image by thresholding 3x3 neighborhood of each pixel with the center value and considering the result as a binary number as depicted in Fig.6. Later histogram of the labels is used as a texture descriptor.

Fig.6 Basic LBP operator for 8 sampling points on a circle of radius 1.
A histogram of the labeled image I(a, b) can be defined as below, where "m" is the number of different labels produced by the LBP operator. This histogram contains information about the distribution of the local micro patterns, such as edges, spots and flat areas, over the whole image. In this paper, the proposed method applies LBP operator for all 40 GMPs and respective histogram are concatenated to build a global description of the face.

IMAGE COMPRESSION USING DISCRETE COSINE TRANSFORMS
This section explains briefly about image compression using Discrete Cosine Transform (DCT) technique. The DCT helps separate the image into parts (or spectral sub-bands) of differing importance with respect to the image's visual quality. The general equation for a 2D DCT of an N x M image is defined by the following equation.
First 8 x 8 block of an image is multiplied by "S" to transform rows and then it is multiplied by "S`" to transform columns. This results in a block matrix with 64 DCT coefficients Cpq where p and q range from (0, 7), in which the top-left coefficient c00 correlates to low frequencies of original image block and c00 corresponds to higher frequencies. It must be noted that human eye is most sensitive to low frequency components.
Block coefficient matrix which is derived from previous step is compressed using quantization. Varying level of image compression can be chosen using specific quantization matrices depending on the intended image quality intended for future processing. Hence there will be a trade-off between quality and compression to suit different needs. There exists a standard image compression matrix "Q50" which considers human visual system and helps to achieve quality level of 50 and high compression with excellent decompression quality as well.
Quantization is achieved by dividing each element in the block coefficient matrix by corresponding element in the Q50 matrix and rounding off the results to the nearest integer value. In this course of division, many of the coefficients become zeroes. Also coefficients situated near the upper left corner corresponds to lower frequencies to which human eye is most sensitive and can be retrieved and lower right coefficients which includes most of the zeroes corresponds higher frequencies which can be discarded. Also only non-zero coefficients are used for image reconstruction. (14) In the final step of compression, all coefficients of "R" are encoded to a stream of binary data. It must be noted that after quantization most of the coefficients will become zero and hence JPEG takes advantage of this fact and follows N o v 1 5 , 2 0 1 3 encoding in zig-zag manner as depicted in the Fig.7. Decompression is carried out by multiplying the corresponding element of quantized matrix "R" with quantization matrix Q50.

IMAGE COMPRESSION
The main aim of this method is to achieve performance improvement of Local Gabor Binary Pattern Histogram Sequence method in terms of recognition accuracy. To achieve this, image preprocessing is also considered as the main factor. The decorrelation of local features using DCT helps in recognition of faces with expression and pose variations accurately. The secondary aim of this method is to alleviate the problems caused by vast datasets. To achieve this, train dataset is stored in compressed domain. This helps in reducing the storage space and avoids image splicing and preserve the privacy of datasets. Detailed flowchart of the proposed method is depicted in Fig.8

A. Preparation of input data using preprocessing steps
The preprocessing is a very essential step in image processing. Many literatures and recognition methods focus more on actual recognition algorithms rather than cleaning up the input data such that the expected result matches the actual result thereby improving the accuracy. In this paper the experimentation is carried out after careful preprocessing of the images and a clear difference is observed with accuracy of face recognition results. Data preparation is carried out in two steps. They are cropping and illumination normalization.
The images are cropped according to the provided eye positions, resized to 50 x 80, succeeded by illumination normalization. The reason behind taking 50x80 [width x height] sized images is that face has more valuable information in vertical direction. Also with resized 50x80 images, memory required for storage and processing time will be reduced considerably which leads to performance improvement to face recognition system.
Main aim of image cropping step is to extract more information from face images which actually add importance and enhances the accuracy of face recognition. In dynamic handing of real time systems, some typical situations occur such that, the distance between camera and person will not be uniform always. This may lead to capturing of face image of a person along with hairs, head, neck, shoulders, hands, etc. It is essential to remove this information that fetches less or no value for actual recognition but may mislead the system. Hence, cropping is an essential step which helps to retain only necessary image information. The steps involved in cropping are as follows.
The Canny edge detector [15] is applied to the original face image of size 256x384 and gray-scale image is converted into binary image. In binary images edges are clearly visible.
Horizontal and vertical projections of binary image are taken. N o v 1 5 , 2 0 1 3 The maximum bins in horizontal and vertical projection are calculated. This indicates the location of eyes and aids in marking the same.
Once eye position is marked, width of the face is calculated as the face is symmetrical with respect to eye positions.
Finally, a border of face is marked with the help of width of face, eye location and edges in binary image.
This cropped face image is resized to 50 x 80. This section explains details of illumination normalization based on histogram of an image and scaling function. In order to develop adaptive contrast enhancement method which suits for all kinds of images of differing illumination levels, we studied the histogram of couple of images (considerably more images with different illumination variations) and concluded that histogram is not concentrated uniformly throughout. As illumination varies, histogram distribution in each bin changes and it is concentrated at particular left, right and corner regions as shown in Fig.10. Analyzing this, we propose a novel enhancement algorithm and it is providing promising result for low, medium and high contrast images. It balances the images in contrast and illumination variations with improved and enhancing edges and preserving the same. This leads to explore more information in enhanced images and helps in improvement of face recognition accuracy.

Fig.10 Study of histograms.
The steps involved for the proposed method are as follows.
Histogram for cropped image is taken and its maximum value is calculated and it assigned as max1. Second maximum value is assigned as max2 and threshold is set equal to max2 / 4. The values which are greater than or equal to threshold in particular region of interest are counted, and is set as C1, C2, C3, C4 corresponding to different regions.
Maximum value among count values of previous step is identified and set as Cmax=max (C1, C2, C3, C4).
It so happens that there can be two count values which are very near or equal. In such case, higher range is selected.
Based on maximum count value and particular range, the limit for histogram is calculated by searching for histogram values less than two on left and right side of histogram. The left minimum value and right minimum values are denoted by lmin and lmax which represent the intensity values. This acts as a terminator on both the sides to set lmin and lmax.
Values which are below intensity lmin in the original image are set to lmin and to intensity lmax which are greater than lmax in the original image.
Finally scaling function is applied, which scales function from 1 to 256 to create missing intensity values.
Histogram of an image after applying proposed above algorithm may not have all the intensity values in it. In order to enhance and make all the intensity levels to be present which makes uniform variations in illumination, we use scaling function. The scaling function scales given range of input from 1 to 256. The algorithm is as follows and the sample preprocessed face image is shown in Fig. 11.
The maximum and minimum value present in an image is calculated.
If the difference between maximum and minimum value is zero, then pixel values of image is assigned to zero. N o v 1 5 , 2 0 1 3 If not, subtract minimum value from each pixel value in an image and divide by the difference of maximum and minimum value and finally multiply by 255 and the result of is rounded off.
The above steps are repeated for all pixel values in an image.

B. Storing preprocessed datasets in compressed domain
In most of the face recognition/verification applications, problems caused by vast datasets are still an open challenge. This memory constraint is solved by storing train and test images in compressed domain. During runtime decompression is applied to the compressed images. The preprocessed images are compressed using 2D-DCT and encoded using Q50 quantizer from JPEG standard as shown in the Fig.12. The non-zero significant DCT coefficients which are less in number are stored in the database which leads to huge reduction in storage space. The fact that face images carry information in lower frequency band helps in making most of the DCT coefficients zero. Exploiting this factor aids to achieve high compression and to save memory. The computation of feature vector is performed by decompressing stored DCT coefficients followed by de-quantization and reconstruction of train images as shown in Fig.13. As and when new images added up with train dataset, the process of compression and decompression can be adopted as an automated process.

Fig. 13 Decompression of DCT coefficients
The pre-processed face image as shown in the Fig.14 is divided into 8x8 blocks and only non-zero coefficients are saved out of all coefficients in proper order as depicted in Fig.15 and 16. At the receiving end, with the help of dequantization followed IDCT, original image is reconstructed and no noticeable losses are observed.

Fig. 16 Non-zero coefficients retained.
High compression is achieved by applying DCT followed by Q50 quantizer as most of the DCT coefficients zero. The same is explained in the in the form of blocks of DCT coefficients in Fig.15. Depending on the information present in an image block, it is essential to select non-zero DCT coefficients by ignoring zero coefficients. At the time of decompression a matrix is formed based on this non-zero coefficients and making remaining places filled with zeros. Then dequantization is applied followed by IDCT to reconstruct an image block.

C. Application of DCT to local features
During the runtime, the Gabor operator is applied for decompressed train images and Gabor Magnitude Pictures (GMP) are constructed. Successively Local Binary Pattern is applied on GMPs and local histograms are formed. The conceptual definition for the proposed idea lies in the fact that the DC component of histograms of local Gabor binary patterns constitutes low frequency components which will not help in recognition and also when these histograms are concatenated, it becomes difficult to differentiate and segregate actual frequency variations which add value for accurate recognition. There exists a high correlation in between these histograms. This high correlation affects the recognition accuracy, hence de-correlation is achieved with the help of DCT application for individual histogram bins which aids in identification of actual frequency variations and highlights the changes in between two histograms thus improving the recognition accuracy. This also exploits full utilization of Gabor filtering and face recognition among faces which has expression variations is accurate. And also one important point is to be noted that if recognition is carried out with application of DCT after concatenation of histograms hits accuracy of recognition, because there is high a chance of mixing of various frequencies from the Gabor transform.
This section explains DCT coefficients used for decorrelation of local features. DCT helps to separate the image into parts (or spectral sub-bands) of differing importance with respect to the image's visual quality [16]. The general equation for a 1D DCT of an NxM image is defined by the following equations.
After application of 1D-DCT, the left coefficient c0 corresponds to low frequencies of original vector and right coefficient c7 corresponds to higher frequencies. It must be noted that human eye is most sensitive to low frequency components. In the case of face images, low frequency components are same for all images, hence these does not add any value for matching. After applying DCT for individual histograms, they are rounded using R=round [DCT(p)] and concatenated and then it is stored as a single feature vector. The variation in frequency leads to better result in face recognition system.
Finally histogram intersection is calculated and it is used as similarity measurement for the proposed method. This similarity measure is chosen because comparison of proposed method with the state-of-art method will be straightforward. The common part of histograms is calculated using the following formula, where H G and H P are local feature histograms of gallery and probe images.
The sample preprocessed face image from FERET database and corresponding retrieved identity which is having a different expression from test database is shown in Fig.17. Fig. 18 shows the state of art method in which local feature histograms are concatenated and in turn histogram intersection is calculated as similarity measure in Fig.19. In this case disadvantage is highly correlated histograms may not lead to good recognition accuracy since utilization of Gabor with different scales and orientation are complete.
Hence in the proposed method, to alleviate this problem, DCT is applied to individual histogram bins as shown in the Fig.20 and histogram intersection is calculated as similarity measure as depicted in Fig.21. An alternative method is also proposed in which DCT is applied after concatenation of local feature histogram as shown in Fig. 22 and 23. The advantage of this method is, clear differentiation among the train and test histograms which are decorrelated by DCT is achieved and this leads to an accurate recognition even in case of facial expression changes. In both the cases, recognition accuracy is improved because actual frequency variations are identified and changes in between two histograms are prominent.

A. Face Variations in terms of appearance
Face variations includes various aspects such as variations with respect to pose, alignment, expression, variations with respect to ethnicity, variations with respect to illumination variation, variations with respect to gender and aging and so on. A detailed study on face variations is carried out for the sake of our experimentation which helped extensively for the analysis of feature histograms. Faces, in general can be classified into two broad categories, that is, static faces with no pose variations and dynamic poses where face alignment and illumination variation exists. In the static faces, there are many variations with respect to ethnicity, gender and aging. The variations of faces according to different races can be categorized using mathematical proportions. One of the ratios defined by ancient Greeks is the concept of "phi," the golden ratio of 1:1.618. Phi is the unique point on a line that divides the line into two lines in such a manner that the ratio N o v 1 5 , 2 0 1 3 of the smaller portion to the larger portion is the same as the ratio of the larger portion to the whole line. Different geographic races with particular genetic similarities are Europeans, Eastern Indians, Asians, American Indians, Africans, Melanesians, Micronesians, Polynesians, Australians and Aborigines. But these races can be accommodated into three basic groups as Caucasians, Asians and Africans. Based on the anatomic structural differences and phenotypic characteristic, certain faces will have a broader nasal base, decreased nasal projection, protrusion of upper and lower jaws, bulging of eyes -orbital proptosis, increased soft tissue of the mid face, prominent lips, and increased facial convexity, faces with a high / low dorsumbridge of the nose feature, distance between the eyes can be small/medium/large, shape of the chin, size of the lips and so on.
When face recognition is carried out with gallery images which are taken years apart from probe images, aging factors are required to be considered. Due to several physiologic changes that occur with time, the collagen in the face becomes less elastic and allows the soft tissue features of the face to sag inferiorly. In addition facial fat is generally lost with age making the facial soft tissues thinner. Finally the facial soft tissues gradually wither making the facial bones appear more prominent. Hence the sagging of cheek results in cheeks, tissues around eyes, eye lids, drift of tissues on forehead, wrinkled around eyebrows, elongated nose and so on.
Along with natural face variations, manmade appearance changes to faces for their disguise also need to be considered. For example face with mustache, beard, clean face without mustache / beard, faces with different hair styles like loose hairs, tied hairs, balded, with / without eye glasses, variations due to eyebrow shapes, variations due to hat / caps and so on.

B. Experimental dataset details
This section presents the experimental dataset details and performance measures used to evaluate the feasibility of the proposed method. FERET database is selected for the experiments to ensure that our results are comparable to other existing literatures. All the images are cropped to 50x80 according to the provided eye positions succeeded by histogram equalization preprocessing. Three databases are formed from FERET database, BioID database and MIT-CBCL-facerec database and labeled as DB1, DB2 and DB3 respectively. Details of FERET database is depicted in the Fig.26. Sample face images from BioID and MIT database are showed in Fig.27 and 28 respectively. N o v 1 5 , 2 0 1 3

C. Performance modeling of face recognition
Performance of face recognition is dependent on both internal and external factors. Internal factors include the actual algorithm used for recognition and train image dataset. External factors include variation of test image dataset in terms of illumination, pose, expression and so on. In a face recognition application, the function of system is to map a query image to a label that represents its identity as depicted in Fig.29. [18]. The gallery dataset is denoted as consisting of "x" gallery images whose identity is known to the face recognition algorithm. The query set is denoted as consisting of "y" query images whose identity is unknown to the face recognition algorithm. A face recognition algorithm measures the similarity between query images and each gallery image. The measure which is used in calculating the similarity in between query image and the probe image plays an important role since it connects both internal and external factors of an recognition system [18]. The similarity score is denoted by for the comparison between the query image and the gallery image . For a perfect match, similarity score should be potentially larger quantitatively. N o v 1 5 , 2 0 1 3 where is the index of the gallery image corresponding to the x th sorted similarity score. For each query image it is assumed that there is one and only one gallery image from the same person and hence the largest similarity score is termed as "matching" score and the remaining similarity scores are termed as "non-matching" scores [18].
A term called "perfect recognition" is introduced for performance modeling of proposed face recognition method [18]. The gallery dataset is duplicated to produce the query dataset. The duplicated set is used for recognition and similarity scores of the each query image are obtained. Such similarity scores are called "Perfect Recognition" (PR) [18]. Quantitative characterization of the difference between matching and non-matching similarity scores for data , a performance metric "P" is defined as where, is the matching score, and are mean and standard deviation of non-matching scores of . The mean of "P" is defined as For the quantitative demonstration, the following experimental steps are carried out. The internal factor being the face recognition algorithm is varied and " " is calculated using only gallery data. As a second step, actual recognition rate " " is calculated using different recognition algorithms. It shows that the system with better recognition accuracy has a larger difference between matching and non-matching similarity scores, thus corresponding " " and " " are larger. The value of " " is bigger with proposed method when compared with that of state of art methods thus concluding the accuracy improvement.
For the experimental purpose, 500 images from BioID database are selected randomly and gallery set is formed with neutral expression and the partially occluded images are used as the probes. From MIT-CBCL-facerec database, 200 images per 10 subjects are taken and randomly divided into gallery and probe images. From FERET database -fb set, 800 faces with expression variation and from fc set, 500 faces with lighting variation are selected. The outcome of the experiments shows that proposed method remarkably highlights the robustness and high discriminating capacity of face features and improved performance. Recognition results are tabulated in Table II in comparison with existing state-of-art methods available in literature. The recognition performance for the proposed method is modeled by the standard error rate and recognition rate commonly used in the face recognition field. False Acceptance is defined as an incorrect prediction that recognizes an incorrect match as a correct match. The false negative, on the other hand, is incorrect prediction that recognizes a correct match as an incorrect match. The plots of FAR, FRR and HTER are shown in Fig.30 and 31 respectively.  Another performance measure taken into account is the storage space for train and test databases and processing time for the recognition. It is being taken care that all the images are preprocessed and cropped to 50x80 such that it occupies only 1/4 th space that of original 128x128 sized images. If the database is stored in compressed domain storage space reduces 1/3 rd that of uncompressed images storage space and processing time is reduced by a factor of 3. The performance plots of storage and processing time is depicted in Fig.32 and 33 respectively. N o v 1 5 , 2 0 1 3

Conclusion
This paper proposes a novel method of face recognition using de-correlation of local features using Discrete Cosine Transforms (DCT). The DC component of histograms of local Gabor binary patterns which constitutes low frequency components which will not help in actual recognition are de-correlated using DCT technique. This differentiation and segregation of actual frequency variations adds value for accurate recognition. This method employs a non-statistical procedure which avoids training step for face samples thereby avoiding generalizability problem which is caused due to statistical learning procedure. Face image preprocessing steps are given more stress to achieve recognition accuracy. The performance modeling is carried out by varying both internal and external factors of face recognition system and improvement is shown considerably high in terms of recognition accuracy and reduction in storage space by storing train images in compressed domain. All the images are preprocessed and cropped to 50x80 such that it occupies only 1/4th space that of original 128x128 sized images and the database is stored in compressed domain such that storage space reduces 1/3rd that of uncompressed images storage space and processing time is reduced by a factor of 3.