A Feature Vector Compression Approach for Face Recognition using Convolution and DWT

The biometric identification of a person using face trait is more efficient compared to other traits as the co-operation of a person is not required. In this paper, we propose a feature vector compression approach for face recognition using convolution and DWT.The one level DWT is applied on face images and considered only LL band. The normalized technique is applied on LL sub band to reduce high value coefficients into lower range of values ranging between Zero and one. The novel concept of linear convolution is applied on original image and LL band matrix to enhance quality of face images to obtain unique features. The Gaussian filter is applied on the output of convolution block to reduce high frequency components to generate fine-tuned feature vectors. The numbers of feature vectors of many samples of single person are converted into a single vector which reduces number of features of each person. The Euclidean distance is used to compare test image features with features of database persons to compute performance parameters. It is observed that the performance recognition rate is high compared to existing techniques


INTRODUCTION
The physiological and behavioral characteristics of a person are used to recognize a living person which leads to Biometric system. The physiological traits are almost constant throughout the life time of a person and are fingerprint, palm print, Iris, DNA etc. The behavioral traits are varying based on mood, circumstances and environment around a person and examples are voice, signature, gait, keystroke etc. The biometric systems are broadly classified into verification system and identification system. In verifications system, the person's identity is declared by comparing features of a person with earlier stored features in the system which is one to one matching, i.e., the claim is accepted or rejected based on preconditions. In identification system, the person's identity is declared by comparing features of a person with earlier stored features of many persons i.e., one to many matching. The identification system is computationally expensive and complex, whereas, the verification system is simple, less expensive and complex. The general biometric system has enrollment section, test section and matching sections. The biometric traits of persons are captured using sensors. The captured images are preprocessed and features are extracted to create database in the enrollment section.
The biometric trait to be tested is captured using sensor, preprocessed and features are extracted to create single feature vector or matrix in the test section. The matching section is used to compare test biometric trait features with database biometric trait features to identify a person based on distance formulae or classifiers. The advantages of biometric identification system compared to traditional methods such as personnel identification number (PIN), identity card, smart card etc., are the biometric traits can't be lost as traits are attached to a person.The face recognition is used in online image search, surveillance in crowd, identification of terrorist in a mob, entry into corporate offices etc. The challenges in face recognition are variations in the light intensity, pose variations, hair occlusions, blinked eye etc.
Novel aspects of this proposed method are: The LL-band coefficients are normalized to convert high coefficient values between Zero and one.
The unique features are generated by convolving original face image with normalized LL-band matrix.
The Gaussian filter is applied on convolution output to reduce high frequency noise components.
Many feature vectors of single person are converted into single unique feature vector by compression to reduce more number of samples per person to one sample person.

LITERATURE REVIEW:
In this section, the existing techniques of face recognition are discussed.
Ahilandeswari et al., [1] proposed multimodal biometric system based on facial, fingerprint and speech. The Eigen faces are used for face identification, minutia features are used for fingerprint matching and cepstral analysis is used for speech identification. The features of all three traits are fused at matching level to enhance performance of architecture.Jianwn wan et al., [2] proposed a novel cost sensitive semi supervised discriminate analysis called PCSDA. Using a simple L2 approach to predict the label of unlabeled data and then learns the projection direction by incorporating costs into both labeled data andunlabeled data comparing wits CS3DA, PCCDA approaches. Using L2 approach to predict the label of unlabeled face images is more accurate and robust than the spares representation. Using CS3DA method approximates the pair wise Bayesian risk only when the classes are balanced and without outliers in data base sets. Hailing Zhou et al., [3] proposed recent advances on face recognition. 3-D data includes facial geometry information, increasing robustness to viewpoints and elimination variation compared with usual images. 3-D face recognition can achieve better recognition accuracy then the usual face recognition, although 3-D data are insensitive to illumination variations, it is still difficult to recognize faces in the options of visible light. Multimodal modalities can achieve better performance than a single modality. IR images are acquired for 2-D imaging sensors, IR face recognition also suffer from sensibility to pose variations. N o v e m b e r 01, 2 0 1 5 Rupali L Telgad et al., [4] proposed biometric system based on facial and fingerprint. The Euclidean distance matches are used for face and fingerprint identification. Minutiae and Gabor filter features are used for fingerprint recognition. Principal component Analysis (PCA)is used for face identification and dimensionality reduction. The feature of two traits approach provides a good result. The recognition rate is increased and the error rate is decreased with the help of two traits. Madeena Sultana et al., [5] proposed illumination insensitiveness of the sub bands of Dual-Tree Complex Wavelet Transform (DTCWT) based on different scales. Extensive illumination variations produces high recognition rate even with a single sample using novel face recognition system approach and weighted fusion of low and high frequency sub bands are used for feature extraction. To identify the adoptive weights during uncertain illumination conditions, a novel fussy weighting scheme are used and adaptive normalization approach is applied for illumination quality announcement of the poor lit samples while retaining the good quality samples. Jiwen Lu et al., [6] proposed multiple statistic features and localized multi-feature metric learning method for a new image set based face recognition identification. Two kernel based metric learning algorithms called localized multi-kernel metric learning and localized multi-kernel multi-metric learning are used for extracting effectively combined multiple statistic features from face image set. Efficient kernel approximation methods are used to improve the kernel estimation speed for specific combined statistic features.
original LBP pattern. The performance of CLBP features are classified using Support Vector Machine classifier.XuXiaona et al., [8] Proposed a kernel based feature fusion algorithm is applied on multimodal recognition system based on ear and face. The kernel principal component analysis (KPCA) algorithm is used for multimodal recognition based on ear and profile face performs better than ear or profile face for unimodal biometric identification. Sushama S patil et al., [9] proposed fingerprint image enhancement technique s, Basically there are two types of fingerprint identification system, automatic fingerprint Authentication system and automatic finger print verification/identification system are uses for recent advances in fingerprint image enhancement to change in finger position, finger condition and finger pressure.
SheetalChaudhary et al., [10] Proposed a multimodal biometric recognition system based on fusion of palm print, fingerprint and face. The matching score level is carried out for fusion of three biometric traits based on the proximity of feature vector and template.Jian yang et al., [11] proposed sparse representation classifiers steered discriminative projection with application to face recognition. Sparse representation based classifier (SRC) was used to direct the design of a dimensionality reduction method SRC -DP, in which the SRC achieved better performance and become more efficient. Meng yang et al., [12] proposed robust Kernel representation model with statistical local feature is used for face identification for different conditions, including variations of illumination, expression misalignment and pose. Javier Galbally et al., [13] proposed multiple biometric systems to detect different types of fraudulent access attempts using software based fake detection method. AarohiVova et al., [14] proposed Support Vector Machine (SVM) based fusion of match scores are used for face and fingerprintbiometric traits. The statistical analysis of different kernel methods is used for training SVM. The polynomial kernel radial basis function kernel and multiplier perceptron kernel are used for face and fingerprint identification. Dinakardas et al., [15] proposed principal component analysis is used to extract the features of the fingerprint and iris images.For face image fisher faces are used. The method of minutiae extraction for fingerprint and LBF feature for iris image is used. Jossy P. George et al., [16] proposed transforms domain fingerprint identification based on DTCWT. The test image features are compared with database images using Euclidean distance. N o v e m b e r 01, 2 0 1 5

PROPOSED MODEL:
In this section, the new concepts of convolution of original face image with DWT of original face image along with compression of database feature vectors are introduced. The block diagram of proposed model is shown in Figure 1.

Face Database
The standard face databases such as ORL, JAFFE,Yaleand Indian male and Indian female are considered to test the proposed model.     [21]:The transformation is used in an image processing to remove noise effectively and also compress an image. The filter bank combinations of low pass and high pass filters are used on rows and columns of an image to derive one approximation band and three detailed bands. The approximation band is an output of low pass filter and has significant information of an image. The detailed bands are outputs of high pass and combinations of high pass-low pass filters and have detailed information such as diagonal, horizontal and vertical information of an image. The DWT decomposes an image into four sub bands such as approximation band, vertical band, horizontal band and diagonal band in each level. The maximum number of decomposition levels equal to log 2 for N×N image size. The rows of images are passing through low pass and high pass filter to generate corresponding low and high frequency coefficients. The columns of an image are passing through low pass and high pass filter along with rows of low pass filter to generate approximation and vertical bands. The columns again passed through low pass and high pass filter along with output of row high pass filter to generate horizontal and diagonal bands. The LL band has significant information of an original image. The LH, HL and HH bands has vertical, horizontal and diagonal information of an original image. The original image can be reconstructed by considering only LL band image and omitting other insignificant information from other bands. The DWT is applied on face databases such as JAFEE, ORL, and Yale, Indian male and Indian female of resized dimensions of 256×256. The LL band is considered for further processing as it has significant information of face images and has less noise component.

Normalization:
Normalization used in the proposed method is to convert high values of LL band coefficients into a range of moderate values. The normalization is applied on LL sub-band coefficients of an image to convert high coefficients values into lower values. Each LL coefficientvalues are divided by maximum coefficient value to convert LL coefficient values range from 0 to 1, as given in equation (1).
The advantage of normalization is the number of bits required to represent each LL coefficient reduces from more than eight bits to less than eight bits. The real time system complexity and memory reduces, whereas speed of computation increases.
The linear convolution matrix has unique coefficient values compared to original matrix coefficient values; hence the concept of convolution is used for better classification of images.

Gaussian Filter:
It is applied on output of convolution block to enhance further the quality of input face images by removing high frequency edges and helps in improving the matching accuracy.An example of Gaussian filter on convolution output matrix is as follows. The Gaussian filter mask matrix is applied on convolution output matrix is given in matrices 7 and 8.
The procedure for Gaussian filter output is as follows.
1. The Gaussian filter mask (h (x,y)) is generated using Gaussian mask coordinators and standard deviation as given in matrix 9 and equation 10.
Gaussian mask coordinators =  The 3x3 Gaussian mask is shifted right and down on X to compute other coefficient value of Gaussian filter output given in matrix 17.

Compression of feature vectors:
The more number of sample images of single person are converted into single sample per person using compression thesix column vector features of sixface images of single person, say 1_1, 2_1, 3_1, 4_1, 5_1and 6_1, are converted into single feature column vector by taking average of six column features as shown in fig 9. The advantages of converting six columns into one column are (i) Time to compare test image with database images reduces (ii) Memory requirement in real time system reduces (iii) The features of single column are more effective compared to six column vector features .
One column feature vector

Euclidean distance:
The distance between database features (pi) and test features (qi) is given in equation (18).
Pi=Coefficient values of vectors in the database.

Qi=Coefficient values of vectors in the test image.
The ED is used to find similarities and dissimilarities among face images to test performance of biometric system.

PERFORMANCE ANALYSIS OF PROPOSED METHOD:
In this section, the definitions of performance parameters and experimental results of various techniques are discussed.

4.2
Performance Analysis: The performance parameters such as FRR, FAR, TSR, EER and Max. TSR are computed and compared for different face databases andtechniques to verify the efficiency of biometric system.

4.2.2Analysis using ORL face database:
The database is created by considering 30 persons with 6 samples per person i.e., 180 samples inside the database and the ninth image of every person is considered as test face image. Ten persons are considered as outside database with one sample per person.The inside and outside databases are used to compute errors such as FRR and FAR respectively. The variations of FAR, FRR and TSR with respect to threshold are shown in figure 11,

Comparison of Proposed Method with existing Methods:
The percentage TSR of proposed method for ORL database is compared with existing algorithms presented by Swarup Kumar DandpatandSukadevMeher [22],Pallavi D. Wadakar and MeghaWankhade [23], and Murugan et al., [24] andAjay et al., [25].is given in table 5. It is observed that the percentage TSR is high in the case of proposed method compared to existing algorithms. The performance of proposed method is better compared to existing algorithms for the following reasons. (i) Normalization performed on 2D DWT reduces the magnitude values of coefficients. ii). unique features are obtained using convolution. iii). Removal of high frequency noise components which help in matching accuracy using Gaussian filter .iv).Compression of Database features storing in to single column vector improves the matching speed.

CONCLUSION:
The identification of a person using physiological trait face is more efficient compared to traditional methods of recognition.
In this paper feature vector compression approach for face recognition using convolution and DWT is proposed. The one level DWT is applied on face images to compress and remove noise in the face images. The LL band is considered and normalized to convert range of high coefficient values to less than one. The convolution is applied between face images and normalized LL band matrix to get better features of faces images. The Gaussian filter is applied on convolution features to enhance further quality of input images. The concept of converting many feature vectors of single person converted into single vector which reduces number of vectors and number of features of a single person. The Euclidean N o v e m b e r 01, 2 0 1 5 distance images to compute performance parameters. It is observed that the performance of the proposed algorithm is better compared to existing algorithm. In feature the convolution can be used in the place of Euclidean distance for the matching.