Procedia Computer Science 93 ( 2016 ) 344 – 350
1877-0509 © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the Organizing Committee of ICACC 2016 doi: 10.1016/j.procs.2016.07.219
ScienceDirect
*
Corresponding author .Tel :+91-9740715006 E-mail address : [email protected]
6th International Conference On Advances In Computing & Communications, ICACC 2016, 6-8
September 2016, Cochin, India
Daubechives Wavelet Based Face Recognition Using Modified
LBP
Shivakumar Dalali
a,*, Suresh L
ba Dept. of CSE, Cambridge Institute of Technology,Bangalore-560036,India. b, Dept. of CSE, Cambridge Institute of Technology,Bangalore-560036,India.
Abstract: With vast and fast development of information technology in our daily life plays an important role. Storing significant information is a big challenge for researchers. Face recognition is one of the most important fields in information technology where storing minimum significant information and accurate recognition is most challenging. This paper proposes an idea of Daubechives wavelet based with modified Local Binary Pattern (LBP) for face recognition with minimum significant information. Applying various Daubechive wavelets with single level decomposition, by considering only the approximation wavelets co-efficient reduces the storage capacity and gives significant information. Then modified LBP is applied on significant wavelet coefficient for simple, fast and accurate face recognition. Proposed work shows higher percentage of face recognition rate by using lesser features with effective threshold.
© 2016 The Authors. Published by Elsevier B.V.
Peer-review under responsibility of the Organizing Committee of ICACC 2016.
Keywords: Face recognition; Daubechives wavelet; Local Binary Pattern
1.Introduction
Biometric is a study of recognising human based on physical or behavioural traits. Among different biometric traits, face recognition receive great amount of attention in present scenario. Extracting discriminate information from an image is one of the key components, for biometric system particularly in face recognition. The commonly used algorithms are principal component analysis (PCA), Gabor phase encoding and Local Binary Pattern (LBP) for feature extraction. The LBP based method has more advantage in face recognition1. LBP was
originally proposed as a texture descriptor with simple and fast to compute 2,3.
PCA: Principal component analysis is a multivariate technique where analyses of face data are described by several inter correlated dependent variables. The goal of PCA is to reduce the dimensionality of data while retaining the variation present in the dataset. The key idea of the PCA method is to find the vector that best © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
account for the distribution of face images within the entire m×n image space. PCA subspace is defined by the Eigen vectors of the covariance matrix and in this subspace 8,11.
Gabor phase encoding: The Gabor filter is used to simultaneously capture spatial and frequency information to overcome a signal represented either in the time or frequency domain. A filter capable of simultaneously capturing spatial and frequency information plays an important role in numerous systems as a feature extractor. Magnitude, phase and orientation are the three basic features produced by Gabor filter. Face recognition based on phase encoding plays important role. Gabor phase encoding used in face recognition suffers with drawback of storing huge amount of data to analyse 12,15,16,17.
1.1 Local binary pattern: LBP is one of the old and most popular local feature based method proposed by Ojala et al. It is basically used for texture description and first time Ahonen et al applied for face recognition6. Later
some authors developed LBP only for recognizing human faces. For example Yang et al3,4 applied LBP for face
recognition with hamming distance constraint. Chen et al5,7 used statically LBP for face recognition18
.
1.2 Discrete wavelet transform: DWT is used in the field of face recognition. The main reasons for its usage are to maintain the significant information in low frequency region, and the great flexibility for choosing its basis functions with low computational complexity. DWT is a signal analysis method of the time scaling and shifting, it’s an advantage for multi-resolution analysis. DWT is a time frequency localization analysis which has the capacity of local features in the time domain and frequency domain. It has fixed window but the shape of its time window and frequency window can change. With wavelets one can analyse low frequency components of the signals with high frequency resolution. Simultaneously looking at high frequency components with high time resolution7,9,10 .
The wavelet transform represents information in terms of the wavelet basis. These are formed through scaling and shifting a single function called the wavelet function. The wavelet transform provides a combination of time and frequency localization. This localization is important for non stationary data. The DWT yields a compact representation of features without a prior knowledge of data statistics13 .
There are two main wavelet functions one Haar wavelet (Db1) an another Daubechives wavelet (DbN). Db1 is the oldest and simplest wavelet function and it compactly supports orthogonal and discontinuous wavelets, which makes an exception from the other wavelet function. DbN are commonly supported orthogonal wavelet, where N denotes number of vanishing moments. Db1 is discontinuous with increasing N and the regularity also increases. Some of DbN’s are less from symmetry; others are very far from symmetry. Db1 have explicit expression where as DbN doesn’t have any explicit expression14.
2.
PROPOSED METHOD
Degradation of an image, different face poses, keeping necessary information in an image are the main challenges for face recognition algorithms. To achieve better recognition rate and the significant feature extraction requires effective pre-processing before face recognition algorithm. Discrete wavelet transform is the one of a major pre-processing method to extract significant information in an image. Daubechive wavelets are used to extract only approximation co-efficient with single level decomposition to reduce the information for face recognition algorithms as shown in the Figure (1)
.
Figure 1. Chart to describe basic decomposition steps for images. 2.1.Modified LBP
LBP operating on each pixel, it requires more computations and memory to store significant features for face recognition [1]. This paper focuses on effectively to reduce computations and storage by modifying existing LBP to operates on approximation DWT coefficients. These coefficients are divided into number of sub blocks of size n*n (e.g. 2x2, 3x3, 4x4, etc), and find the average value of each sub block to form a new matrix. Each entry in the matrix represents the average value of their respective sub-blocks. Consider 3*3 sub matrix to determine the LBP value of each element in the matrix as shown in Figure (2).
2 8 1
2 .
c i A c i iLBP
¦
m A
A
(1) Where m (k) is defined as( )
1;
0
.
0 ;
0
k
m k
k
!
½
®
¾
¯
¿
The LBP value of the central element in the sub matrix is computed by comparing it with its neighbourhoods is given by Eq (1)
Figure 2. Sub-block pixels
Histogram of LBP matrix of size M*N is determined by the Eq (2)
> @
1 1,
,
,
0,
M N i jH b
¦¦
f LBP i j b
b
B
(2)f X Y
,
^
1,0,X YOtherwiseThe binary value of an LBP is defined as the number of bitwise transitions and is calculated by the Eq (3)
5 1 3 7 4 8 2 0 c c c c c c c c V LBP S A A S A A S A A S A A S A A S A A S A A S A A (3)
In practice, the mapping from LBPP, R to
2
,
V P R
LBP
(Superscript “V2” means that the uniform patterns have a Vvalue of at most 2), which has P*(P-1) +3 distinct output values is implemented with a lookup table of 2p
elements.
The dissimilarity of sample and model histograms is a test of goodness-of-fit, which could be measured with nonparametric statistic test. In this study the dissimilarity between a test sample S and a class model T is measured by the Chi-square distance as shown in Eq (4).
2 1 , N n n n n n S T D S T S T
¦
(4) Where N is the number of bins, Sn and Tn are the values of the sample and model images at the nth bin.2.2.Non-overlapping
The new matrix which is an average value of non-overlapping sub blocks of approximation DWT coefficients of size n*n is calculated using Eq (5). The average value of non-overlapping sub blocks of new matrix depicted in Figure (3) which reduces storage. This reduction of storage is one of the solutions for volume characteristics issue in big data and used in face recognition. Non overlapping sub blocks of an average values requires less memory by increasing the size of sub blocks and observed in the results.
.
1 1
1
,
*
n n V i jA
P i j
n n
ª
º
«
»
¬
¦¦
¼
(5)Where P (i, j) is the pixel in the sub block
Figure 3. Non- overlapping sub-blocks
Parallel processing can be used to create a new matrix because of non overlapping sub blocks which in turn helps for big data.
2.3.Hierarchical Multi-scale Modified LBP
The performance of single basic LBP operator is limited. Multi-scale or multi-resolution will have more image feature. In basic LBP features of different scale are extracted first and then the histogram are concatenated to make a long feature. It suffers with huge feature dimension. In basic LBP non-uniform patterns are clustered into one “non-uniform” pattern. By this scheme much information is lost. To extract more useful features from the image LBP is modified to hierarchical multi-scale LBP. The pattern of bigger radius is non-uniform but its
counterpart in a smaller radius is uniform [1]. Proposed method builds multi-scale LBP histogram from big radius to small radius. Firstly consider the LBP map of the biggest radius. Then these pixels are divided into “uniform” to “non-uniform” patterns. Using “uniform” patterns build sub histogram and those with “non-uniform” patterns are further processed to extract their LBP patterns by smaller radius. This process is continued till it reaches the smallest radius or single pixel. Modified LBP gives better accuracy by considering less number of significant features in the face. With this it reduces complexity, time and storage.
3.Experimental Results
3.1 Images without noise
Figure 4 shows that the input images without noise and a person with different positions are considered for modified LBP. The table1 shows that the percentage of recognition rate simulated for 100 different persons with different poses images of size 288×324 from MIT face database without input noise to the images. Recognition rate depends on the size of the non-overlapping sub blocks and also size of the new matrix created for modified LBP. If sub block size is small then percentage of recognition rate is high and threshold value rate is more because of localization. As sub block size increases percentage of recognition rate and threshold value slightly reduces. Using modified LBP methods for fixed neighbourhoods of multi-scale (Ps) and varying radius of the
multi-scale (Rs) and block size gives in an average better percentage of recognition rates.
Figure 4. Input images without noise with different poses.
Table 1. Recognition rate of images without noise using non-overlapping sub blocks.
Non-overlapping sub block size in the input image.
Size of a new matrix to the modified LBP. Modified LBP methods. Threshold value (ranges). % of recognition rate. 2x2 26x21 Ps=8 Rs=4,3,2,1 Block size=10x13 50 -60 99.3 3x3 17x14 Ps=8 Rs=4,3,2,1 Block size=10x13 20-30 99.11 4x4 13x11 Ps=8 Rs=3,2,1 Block size=7x10 10-20 99.01 5x5 10x8 Ps=8 Rs=2,1 Block size=6x8 10-20 99.00
3.2
Images with noise
person with different positions are considered for modified LBP. Table 2 shows that percentage of recognition rate with noisy images slightly reduces by considering the above simulations shown in table1.
Figure 5. Input images with noise and different poses of a person
Table 2. Recognition rate of images with noise using non-overlapping sub blocks.
Non-overlapping sub block size in the input image.
Size of a new matrix to the modified LBP. Modified LBP methods. Threshold value (ranges). % of recognition rate. 2x2 26x21 Ps=8 Rs=4,3,2,1 Block size=10x13 50 -60 98.28 3x3 17x14 Ps=8 Rs=4,3,2,1 Block size=10x13 20-30 98.13 4x4 13x11 Ps=8 Rs=3,2,1 Block size=7x10 10-20 98.05 5x5 10x8 Ps=8 Rs=2,1 Block size=6x8 10-20 97.82 4.CONCLUSION
As the present traditional processing models in the face recognition field uses huge amount of features to analyse, the proposed method uses Daubchive wavelet and modified LBP to overcome the above problem. Daubechive wavelet reduces information into few significant coefficients for less storage. Modified LBP is applied on these significant coefficients to extract very essential features to recognise the face. The experimental results ensure that the idea used in this paper archives simple, fast, accurate face recognition with very less significant data and higher percentage of recognition rate through effective threshold. This method helps in face recognition application like citizen identity cards, voters ID’s and social benefits.
This in turn helps solving the volume issue of big data analysis in biometrics. The future work is applying some more technique to solve the big data characteristics to leverage existing biometrics to big data biometrics. 5.ACKNOWLEDGEMENT
We have used the face images of MIT Face databases for our experiments. Henceforth we would like to thank them a lot.
References
1. Zhenhua Guo, Lei Zhang, David Zhang and Xuanqin Mou “Hierarchical multi-scale LBP for face and palm print recognition”
2. T. Ahonen, A. Hadid, and M. Pietikäinen, "Face description with local binary patterns: application to face recognition," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 28, no. 12, pp.2037-2041, 2006.
3. G. Zhang, Z. Huang, and S. Z. Li, "Boosting local binary pattern (LBP)-based face recognition," SinoBiometrics, pp. 179-186, 2004.
4. Y. Raja, and S. Gong, "Sparse multiscale local binary patterns," Proceeding of 17Th British Machine Vision
Conference, 2006.
5. S. Liao, X. Zhu, Z. Lei, L. Zhang, and S. Z. Li, "Learning multi-scale block local binary patterns for face recognition," International Conference on Biometrics, pp. 828-837, 2007.
6. T. Ojala, M. Pietikainen, and T. Maenpaa, "Multiresolution gray-scale and rotation invariant texture classification with local binary patterns," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, no. 7, pp. 971-987, 2002.
7. Abdallah Mohamed and Roman Yampolskiy “Face Recognition Based on Wavelet Transform and Adaptive Local Binary Pattern”
8. Girish Rao Salanke N S, Dr. Maheswari N, Dr. Andrews Samraj, S.Sadhasivam "Enhancement in the design of Biometric Identification System based on Photoplethysmography data”, Proceedings of 2013. ICGHPC March-2013.
9. Pallavi D.Wadkar, Megha Wankhade “Face recognition using discrete wavelet transforms” IJAET/Vol.III/ Issue I/January-March, 2012/239-242.
10. Wu Fengxiang “Face Recognition Based on Wavelet Transform and Regional Directional Weighted Local Binary Pattern” journal of multimedia, vol. 9, no. 8, August 2014/1017-1023.
11. Patrik Kamencay, Dominik Jelšovka, Martina Zachariasova “The impact of segmentation on face recognition using the principal component analysis (pca)”
12. Angel Serrano, Isaac Martín de Diego, Cristina Conde, et al, “Recent advances in face biometrics with Gabor wavelets: A review”, Pattern Recognition Letters, 2010, 31(5) pp 372-381.
13. Meng, J., Gao,Y., Wang, X., Lin, T., Zhang, J.” Face Recognition Based on Local Binary Patterns with Threshold.” In: IEEE International Conference on Granular Computing, pp. 352-356. San Jose, CA (2010) 14. Mohamed, A.A., Gavrilova, M.L., Yampolskiy, R.V. “Artificial Face Recognition using Wavelet Adaptive LBP with Directional Statistical Features.” In: 12th International Conference on Cyberworlds. Darmstadt, Germany (2012).
15. Iqbal Nouyed and M. Ashraful Amin “Recent Advancements in Gabor Wavelet based Face Recognition” 16. Mounika B.R., Reddy N.J. and Reddy V.B.D. “A neural network based face detection using Gabor filter response” International Journal of Neural Networks ISSN: 2249-2763 & E-ISSN: 2249-2771, Volume 2, Issue 1, 2012, pp.-06-09.
17. N. S. Girish Rao Salanke, N. Maheswari and Andrews Samraj, "An Enhanced Intrinsic Biometric in Identifying People by Photopleythsmography Signal", Proceedings of the ICSIP 2012, Ó Springer India 2013. 18. Mr. Shivakumar Dalali, Dr. Suresh L and Dr. Chandrakanth Naikode “Case study: Leveraging Biometrics to Big data” International Journal of Engineering Research (IJER) Volume 4, Issue 5. May 2015, Pp: 396-398. ISSN: 2319 – 6890.
