A Novel Illumination Invariant Face Recognition Method Based on PCA and WPD Using YCbCr Color Space

(1)

Procedia Computer Science 92 ( 2016 ) 181 – 187

Peer-review under responsibility of the Organizing Committee of ICCC 2016 doi: 10.1016/j.procs.2016.07.344

ScienceDirect

2nd International Conference on Intelligent Computing, Communication & Convergence

(ICCC-2016)

Srikanta Patnaik, Editor in Chief

Conference Organized by Interscience Institute of Management and Technology

Bhubaneswar, Odisha, India

A Novel Illumination invariant Face recognition method based on PCA and

WPD using YCbCr color space

Archana H. Sable

a

, Prof. Sanjay N. Talbar

b

a

School of Computational Sciences,S.R.T.M.University,Nanded and 431606, India

bDepartment of Electronics and Telecommunication Engg., Shri Guru Gobingji Singh Institute of Engg. & Tech.,Nanded(M.S) India,

Abstract

Today, much research on face recognition has focused on using grey-scale images. With the increasing availability of color images, it makes sense to develop approaches for integrating color information into recognition process as the grey-scale approaches is sensitive to lighting variations. In this paper, we have proposed a novel two phase method, i.e., YCbCr-WPD-PCA-Mah. In the first phase, we convert the each training face into Y, Cb and Cr components and then decompose Y, Cb and Cr components into k parts using the Wavelet Packet Decomposition(WPD).finally perform PCA for k-times on Y, Cb and Cr subbands, to get k eigenspaces and k feature vectors for each Y, Cb and Cr subbands.In the second phase i.e., classification phase, the test image is projected onto the Y-eigenspace, Cb-eigenspace, and Cr-eigenspace after being decomposed into k part using WPD.Then, the Mahalanobis Distance is computed between the test image and all the training images in Y-subspace, Cb-subspace, and Cr-subspace.The Mahalanobis Distance is computed between the merged feature vectors.In deceision level, we compute the mean of the Mahalanobis Distances obtained from Y, Cb and Cr subspaces. The face has the best match with the test image is which has the minimum distance. The accuracy of the proposed method i.e., YCbCr-WPD-PCA-Mah has been identified and a comparison was performed in terms of recognition rates or Equal Error Rate(EER) or the Receiver Operating Curves(ROC).

Selection and peer-review under responsibility of scientific committee of Interscience Institute of Management and Technology. Keywords: Face Recognition; YCbCr; WPD; PCA; Mahalanobis distance.

* Corresponding author. Tel.: +91-9765356793; E-mail address:[email protected]

1.Introduction

BIOMETRICS is the science of establishing the identity of an individual based on the physical, chemical or behavioral features of the person. Among the features that may be are: face, fingerprints, hand geometry, handwriting, iris, retinal, vein, and voice.Face recognition system is a computer application for automatically identifying or verifying a person from a digital image or a video frame. A survey of various FR techniques has been provided in Ref. (W. Zhao, R. Chellappa,P. J. Phillips & A. Rosenfeld 2003).

(2)

Performance of an FR system is subjected to variations in imaging conditions like illumination and pose (X. Zou, J. Kittler & K. Messer 2007). Conventional enhancement techniques focus on the problem of uniform illumination variation. The primary task in an FR system is the extraction of the features. Feature extraction can be done by one of the following methods- Generic, Template based, Structural matching and Transform-based (Discrete Cosine Transform (Yuann

Jumei & Pan Hongxia 2011), Discrete Fourier Transform and Discrete Wavelet Transform (DWT) (R. M. Ramadan & R. F. Abdel Kader 2009)). The rest of the paper is organised as follows: Section 2 introduces the wavelet packet decomposition(WPD). Section 3 introduces the Principal Component Analysis(PCA).Section 4 introduces the proposed algorithms i.e., YCbCr-WPD-PCA-Mah. Section 5 shows Results that have been tabulated and are graphically depicted. Section 6 contains the conclusion.

2.Wavelet Packet Decomposition (WPD)

Discrete wavelet transform(DWT) is a well-known signal analysis tool, is widely used in feature extraction, compression and denoising applications. Discrete wavelet transform has been used in various studies in face recognition. In discrete wavelet decomposition, only LL is further decomposed. In case of DWT however, oscillatory patterns are not represented in a proper manner as the wavelets are ill suited. The oscillating variations in the intensity can only be described by small-scale wavelet coefficients. But as, these small-scale coefficients has little energy, and are quantized to zero even at high bit rate. The drawback of DWT is overcomed by the wavelet packet decomposition(WPD) also known as wavelet packets. Wavelet packets are able to represent the high frequency information. In discrete wavelet decomposition, only LL is further decomposed. Conversely, in wavelet packet decomposition all LL, LH, HL and HH are further decomposed as shown in fig. 1(b) and 1(c ). Fig. 1(b) illustrates the wavelet packet decomposition at level 1 of the original image as shown in fig 1(a). However, the wavelet packet decomposition at level 2 is given in fig. 1(c).

(a) (b) (c)

Fig 1. Wavelet Packet Decomposition(a) original image (b) at levels 1 (c ) at levels 2.

Wavelet packets represent a generalization of multiresolution decomposition. In the wavelet packets decomposition, the recursive procedure is applied to the coarse scale approximation along with horizontal detail, vertical detail, and diagonal detail, which leads to a complete binary tree.

3.Principal Component Analysis (PCA)

Principal Component Analysis(PCA) is one of the most popular appearance-based methods used mainly for dimensionality reduction in compression and recognition problems.Principal component analysis(PCA) also known as Karhunen-Loeve expansion, which has two useful properties when used in face recognition. The first is that it can be used to reduce the dimensionality of the feature vectors. The second useful property is that PCA eliminates all the statistical covariance in the transformed feature vectors.Sirovich and Kirby first used PCA to efficiently represent pictures of human faces.They showed that any particular face can be (i)economically represented along the eigenpictures coordinate space, and (ii) approximately reconstructed using just a small collection of eigenpictures and their corresponding projections.Within this context, Turk and Pentland presented the well-known Eigenfaces method fir face recognition in 1991. Since then, PCA has been widely investigated and has become one of the most successful approaches in face recognition.

(3)

PCA is mathematically defined as an orthogonal linear transformation that transforms the data to a new coordinate system such that the greatest variance by any projection of the data comes to lie on the first coordinate (called the first principal component), the second greatest variance on the second coordinate, and so on. PCA is theoretically the optimum transform for given data in least square terms.PCA can be used for dimensionality reduction in a data set by retaining those characteristics of the data set that contribute most to its variance, by keeping lower-order principal components and ignoring higher-order ones. Such low-order components often contain the "most important" aspects of the data. However, depending on the application this may not always be the case .

For a data matrix, XT, with zero empirical mean (the empirical mean of the distribution has been subtracted from the data set), where each row represents a different repetition of the experiment, and each column gives the results from a particular probe, the PCA transformation is given by:

(1) where the matrix Σ is an m-by-n diagonal matrix with nonnegative real numbers on the diagonal and W Σ VT is the singular value decomposition (svd) of X. PCA has the distinction of being the optimal linear transformation for keeping the subspace that has largest variance. This advantage, however, comes at the price of greater computational requirement if compared, for example, to the discrete cosine transform .

4.Proposed method based on PCA and WPD using YCbCr color space (YCbCr –PCA-WPD-Mah):

Much research on face recognition has focused on using grey-scale images. With the increasing availability on color images, it makes sense to develop approaches for integrating color information into the recognition process. Since the grey-scale approach is sensitive to lighting variations, we have chosen an approach which takes into consideration the intensity information as well as the chromatic information such as the YCbCr color space.

4.1. YCbCr Color space:

YCbCr is a family of color spaces used as a part of the color image pipeline in video and digital photography systems.It is a way of encoding RGB information. The RGB color space has been widely used for processing and storing digital image data. This model describes each color as a weighted combination of three base components Red, Green and Blue. However, high correlation between components and mixing luminance with chromaticity makes it very sensitive to changes in imaging conditions such as lighting. Unlike the RGB space, YCbCr separated luminance from chrominances using a linear transform consisting of a weighted sum of the three components. The Y value represents the luminance or brightness component while Cr and Cb values also known as the color difference signals represent the chrominance component of the image.

Y =6 + 0.2567882 R + 0.5041294 G + 0.0979059 B (2)

Cb =128 – 0.148223 R– 0.2909928 G + 0.4392157 B (3) Cr=128 + 0.439216 R – 0.3677884 G - 0.0714274 B (4) Where

Y is luminance, Cb is blue chromaticity, and Cr is red chromaticity.

The combined component i.e., YCbCr representation of the face and the separate components of the YCbCr representation of the face are depicted in the figure.Fig 2(a) shows original face, fig.2(b) shows the YCbCr view of the face, fig.2(c) shows the Y component, fig.2(d) shows the Cb component and fig.2(e) shows the Cr component of the face. From the figure we can say that the person is recognizable from the Y image, but the Cr and Cb components are less recognizable.

Fig.2 . Different views of the face (a) original face (b) YCbCr representation of the face (c) Y component (d) Cb component ( e) Cr component

¦ V W X VT T

(4)

Steps followed in this method are:

Step 1: Convert each face image into Y, Cb and Cr components.

Step 2: Decompose Y, Cb and Cr components obtained from the original images into k parts using the Wavelet Packet Decomposition. In our method, we have used the wavelet packet decomposition at level two. At the second level of decomposition, we obtain one image of approximation(low-resolution image ) and 15 images of details. Therefore, the face image is described by 16 wavelet coefficient matrices, which represent quite a huge amount of information(equal to the size of the input image). Each of the coefficient matrices contains information about the texture of the face.

Step 3: Perform PCA for k-times in the Y, Cb, and Cr subbands(approximations and details). After the training, we get k eigenspaces and k feature vectors for each of Y, Cb and Cr subbands.

Step 4: The k feature vectors obtained from k eigenspaces are merged into one vector. The features are selected corresponding to the largest eigenvalues of the merged and sorted eigenvalues vector.

Step 5: In the classification phase, the test image is projected onto the Y-eigenspace, Cb-eigenspace, and Cr-eigenspace after being decomposed into k part using the Wavelet Packet Decomposition. Then, compute the Mahalanobis distance between the test image and all the training images in Y-subspace, Cb-subspace, and Cr-subspace. The Mahalanobis distance is computed between the merged feature vectors. In the decision level, we compute the mean of the Mahalanobis distances obtained from Y, Cb, and Cr subspaces. The face that has best match with the test image is which has the minimum distance.

Fig.3 . Block diagram of the proposed method.

5.Experimental results:

To evaluate the efficiency of the method discussed, we have used two face databases i.e., face94 database. The face94 database (http://cswww.essex.ac.uk/mv/allfaces/faces94.html) consists of 153 distinct persons. There are 20 images per person, a total of 3060 images. The subjects were sitting at approximately the same distance from the camera and were asked to speak while a sequence of twenty images was taken. The speech was used to introduce moderate and natural facial expression variation. The size 180 by 200 pixels (portrait format).There are three directories: female (20), male (113), malestaff(20).They contains images of male and female subjects in separate directories. None of the 20 samples are identical to each other; so there was no overlap between the training and test sets. Besides, the original images are used in training and testing without pre-processing. Fig 4(a) shows some examples from the face94 database. The FEI face database is a Brazilian face database that contains a set of face images taken between June 2005 and March 2006 at the Artificial Intelligence Lab of FEI, Brazil. There are 14 images of 200 individuals, a total of 2800 images. All images are colorful and taken against a while homogeneous background in an upright frontal position with profile rotation of upto about 180 degrees. The number of male and female subjects is exactly the same and equal to 100.This database is available at (http://www.fei.edu.br/~cet/facedatabase.html). Fig 4(b) shows some examples of image variations from the FEI face database.

(5)

Archana H. Sable, Prof. Sanjay N. Talbar/ Procedia Computer Science00 (2014) 000–000

Fig.4 . (a)Examples of two subjects of face94 database. (b)Some examples of image variations from the FEI face database.

To identify the accuracy of our algorithm, a comparison was performed with other methods in terms of recognition rates. In this experiment, we have selected from face94 database 30 images, 15 images of female and 15 images for male. As there are 20 samples per image, we select 10 samples per image for training and 10 samples per image for testing for part I, so there are 300 images in training set and 300 images in testing set(10 sample per class in the training set). For part II, we have selected from FEI face database 20 images, the images with extensions(-01.-03,-05,-07,-09,-12,-13 and -14) are selected for testing and the remaining images are used in the training.In this experiment, we have compared between PCA, DWT-PCA, WPD-PCA and YCbCr-WPD-PCA-Mah methods in terms of the recognition rates with three , five, and ten samples per class in training set. The results of the tests are summarized in the table 1 . We have seen that the recognition rates increase as we increase the numbers of samples per class. Our method (YCbCr-WPD-PCA-Mah) outperforms other methods; since it reaches recognition rates of 90.50%, 95.87%, and 98.99% with three, five and ten samples per class respectively.

Table 1. Results of the test

We compare the proposed method

YCbCr-WPD- PCA-Mah with the

other methods by plotting Receiver

Operating Curves(ROC)

curves for the different methods with five samples per class which are shown in fig. 5(a) .The experiments are carried out though the use of a laptop(dual CPU T2330 1.4 GHz and 2GB memory) and the MATLAB language as tool of development. The GUI for the application is as shown in the fig.5(b).

Method Number of images Number of Misclassified /rejected images Recognition rate (%) PCA 3 52 67.50 5 30 81.25 10 22 87.11 DWT-PCA 3 64 60.00 5 40 71.87 10 31 80.62 WPD-PCA 3 43 73.31 5 24 85.00 10 17 89.37 YCbCr-WPD-PCA-Mah 3 16 90.50 5 11 95.87 10 6 98.99

(6)

Fig 5 . (a)ROC curves PCA, DWT-PCA, WPD-PCA and YCbCr-WPD-PCA-Mah methods with five samples per class. (b) The GUI for the application

6.Conclusion

In this paper, we evaluate the accuracy of our algorithm the YCbCr-WPD-PCA-Mah, a comparison was performed with other methods via PCA, DWT-PCA, WPD-PCA methods in terms of the recognition rates with three , five, and ten samples per class in training set.We have seen that the recognition rates increase as we increase the numbers of samples per class. Our method (YCbCr-WPD-PCA-Mah) outperforms other methods; since it reaches recognition rates of 90.50%, 95.87%, and 98.99% with three, five and ten samples per class respectively.In the proposed method, we have used the YCbCr color space.The future scope, can be that, we can use other color space and do comparison between them in order to choose the best one. We have also used PCA as dimensionality reduction technique, we can try other techniques such as SVD or FLD.

Acknowledgements

First of all, I thank Lord Shri Gajanan Maharaj for giving me strength and ability to complete this study. Futhermore, I would express my gratitude to my research guide Prof. Sanjay Talbar for their powerful guidance, motivation, valuable suggestions, support and attention throughout this research. I thank all my friends for sharing their experiences and knowledge. Special thank to My Husband for providing moral support and always encouraging me to do good research. Finally, I am very grateful to my parents for their trust everlasting love, care, and indefectible support morally and materially during all my years of studying.

References

[1] Yoo, S.; Park, R.-H. & Sim, D.-G. (2007). Investigation of color spaces for face recognition.Proceedings of IAPR Conference on Machine Vision Application, pp. 106–109, ISBN 978-4-901122-07-8, Tokyo, Japan, May 16-18, 2007

[2]Zhang, Z & Shi, Y. (2009). Skin color detecting unite YCgCb color space with YCgCr colorspace. Proceedings of 2009 International Conference on Image Analysis and Signal Processing, pp. 221–225, ISBN 978-1-4244-3987-4, Taizhou, China, April 11-12,2009

[3]Zheng, W.-S.; Lai, J.-H. & Yuen, P. C. (2005). GA-fisher: A new LDA-based face recognition algorithm with selection of principal components. IEEE Transactions on Systems,Man, and Cybernetics, Part B: Cybernetics, Vol. 35, No. 5, pp. 1065–1078, ISSN 1083-4419.

[4]Jones III, C. F. & Abbott, A. L. (2004). Optimization of color conversion for face recognition.EURASIP Journal on Applied Signal Processing, Vol. 1, No. 4, (October 2003), pp. 522–529, January 2004

[5]Liu, C. & Wechsler, H. (2000). Evolutionary pursuit and its application to face recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 22, No. 6, (August 2002), pp. 570–582, ISSN 0162-8828

[6]Neagoe, V.-E. (2006). An optimum 2D color space for pattern recognition. Proceedings of the2006 International Conference on Image Processing, Computer Vision, & Pattern Recognition, Vol. 2, pp. 526–532, ISBN 0162-8828, Las Vegas, NV, USA, August 26-29, 2006

[7]Phillips, P. J.; Moon, H.; Rizvi, S. A. & Rauss, P. J. (1998). The FERET database and evaluation procedure for face recognition algorithms. Image and Vision Computing,Vol. 16, No. 5, pp. 295–306, ISSN 0262-8856

[8]Phillips, P. J.; Moon, H.; Rizvi, S. A. & Rauss, P. J. (2000). FERET evaluation methodology for face recognition algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 22, No. 10, (August 2002), pp. 1090–1104, ISSN 0162-8828

[9]Rajapakse M.; Tan, J. & Rajapakse, J. (2004). Color channel encoding with NMF for face recognition. Proceedings of 2004 International Conference on Image Processing, Vol. 3,pp. 2007–2010, ISBN 0-7803-8554-3, Singapore, October 24-27, 2004

[10]Sim, T.; Baker, S. & Bast, M. (2003). The CMU pose, illumination and expression database.IEEE Transactions on Patten Analysis and Machine Intelligence, Vol. 25, No. 12,(December 2003), pp. 1615–1618, ISSN 0162-8828

(7)

[11]Torres, L.; Reutter, J. Y. & Lorente, L. (1999). The importance of the color information in face recognition. Proceedings of 1999 International Conference on Image Processing, Vol. 3,pp. 627–631, ISBN 0-7803-5467-2, Kobe, Japan, October 24-28, 1999

[12]Turk, M. & Pentland, A. (1991). Face recognition using eigenfaces. Proceedings of CVPR 1991IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp.586–591, ISBN 0-8186-2148-6, Maui, HI, USA, June 3-6, 1991

[13]Turk, M. & Pentland, A. (1991). Eigenfaces for recognition. International Journal of Cognitive Neuroscience, Vol. 3, No. 1, pp. 71–86, ISSN 0898-929X [14]Wiskott, L.; Fellous, J.-M.; Krueuger, N. & von der Malsburg, C. (1997), Face recognition by elastic bunch graph matching. IEEE Transactions on Pattern

Analysis and Machine Intelligence, Vol. 19, No. 7, (August 2002), pp. 775–779, ISSN 0162-8828

[15]Wyszecki, G. & Stiles, W. S. (2000). Color Science: Concepts and Methods, Quantitative Data and Formulae (2nd). John Wiley & Sons, New York, USA [16] In Zhang, D. & Jain, A. K. (eds.): Advances in Biometrics, Lecture Notes in Computer Science, Vol. 3832, Springer-Verlag Berlin Heidelberg, pp. 64–68,

ISSN 03202-9743

[17]Yang, M.-H. (2002). Kernel eigenfaces vs. kernel Fisherfaces: Face recognition using kernel methods. Proceeding of the 5th IEEE International Conference on Automatic Face and Gesture Recognition, pp. 215–220, ISBN 0-7695-1602-5, Washington, D.C., USA, May 20-21, 2002

[18]Bartlett, M. S.; Movellan, J. R. & Sejnowski, T. J. (2002). Face recognition by independent component analysis. IEEE Transactions on Neural Networks, Vol. 13, No. 6, pp. 1450–1464, ISSN 1045-9227

[19]Belhumeur, P. N.; Hespanha, J. P. & Kriegman, D. J. (1997). Eigenface vs. Fisherfaces:Recognition using class specific linear projection. IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 19, No. 7, (August 2002), pp. 711–720, ISSN0162-8828

[20]Chai, D. & Ngan, K. N. (1999). Face segmentation using skin-color map in videophone applications. IEEE Transactions on Circuits and Systems for Video Technology, Vol. 9,No. 4, (August 2002), pp. 551–564, ISSN 1051-8215

[21]De Dios, J. J. & Garcia, N. (2003). Face detection based on a new color space YCgCr.Proceedings of 2003 International Conference on Image Processing, Vol. 3, pp. 909–912,ISBN 0-7803-7750-8, Barcelona, Spain, September 14-17, 2003

[22]De Dios, J. J. & Garcia, N. (2004). Fast face segmentation in component color space.Proceedings of 2004 International Conference on Image Processing, Vol. 1, pp. 191–194,ISBN 0-7803-8554-3, Singapore, October 24-27, 2004

[23]Etemad, K. & Chellappa, R. (1997). Discriminant analysis for recognition of human faceimages. Proceedings of the First International Conference on Audio- and Video-Based Biometric Person Authentication, pp. 1724–1733, ISBN 3-540-62660-3, Crans Montana, Switzerland, March 12-14, 1997