Documente Academic
Documente Profesional
Documente Cultură
different sets of anthropometric points were used for recognition accuracy. In general, face recognition
asymmetry estimations and the resulting asymmetry algorithms have two major phases; the training stage and
measure was calculated by averaging these estimations. the recognition stage.
Grgic, Delac and Grgic [9] described a database of static
3.1 Training Phase
images of human faces that proposed a tested protocol. A
simple baseline principal component analysis face Training phase represents the first phase where the database
recognition algorithm was tested following the proposed of images for known people is used, and at least one image
protocol, which proving their robustness and efficiency. per known person is available at the database. In this phase
Zhan et al. [10] presented an evolutionary classifier fusion the features for each known face image are extracted and
method inspired by biological evolution is presented to stored in the database. There are three main steps in the
optimize the performance of a face recognition system. training phase that shown in Figure 1, including calculating
Different illumination environments are modeled as a projection matrix from the trained images, then extracting
multiple contexts using unsupervised learning and then the features from the images and finally storing these features to
optimized classifiers are searched. be used in the recognition phase. First step concerns with
Saleh et al. [11] proposed a shift invariant pattern the calculation of the projection matrix, in which Face
recognition mechanism using a feature-sharing recognition algorithm use the training images to calculate
hypercolumn model. In this work a share mape is the projection matrix. For example in Principal Component
constructed to improve the recognition rate and to reduce Analysis (PCA) the projection matrix consists of
the memory requirements of the hypercolumn model. eigenvectors of the covariance matrix, where the covariance
Guoxing et al. [12] a local binary pattern (LPB) has been matrix is derived from the trained images. The second step
proved to be successful for face recognition. A new LPB- deals with the extraction of the features from images. In
based multi-scale illumination preprocessing method was PCA each image is multiplied by the projection matrix to
proposed, in which this method performs better than the formulate the eigenfaces. The eigenfaces represent the
existing LPB-based methods. principal component of the images, or in other words it
Sing, Thakur, Basu, Nasipuri and Kundu [13] proposed a represents the features of the image. Finally the last step in
self adaptive radial basis function neural network based the training phase is to store the features that are extracted
method for high speed recognition of human faces. It has from the trained images [1], [3].
been seen that the variations between the images of a
person, under varying pose, facial expirations, illumination, Calculate Calculate
Store
etc. were quite high. projection feature
feature
Face matrix extraction
Kumar et al. [14] presented a PCA memetic algorithm images as vector in
database
approach for feature selection of face recognition. PCA has input of
training
been extended by memetic algorithm, where the former was phase
used for feature extraction reduction and the later exploited Projection Feature
matrix vector
for feature selection.
Banerjee et al. [15] studied the performance of frequency Figure 1. Training phase of general face recognition
domain correlation filter for face recognition by considering algorithm.
only the phase part of the face images, in which the
dimensions of the phase spectra were reduced by using 3.2 Recognition Phase
frequency domain principal component analysis. Recognition phase represents the second phase, in which
Zhangand and Qiao [16] represented a new gradient Gabor each new unknown face image is analyzed to obtain its
filter to extract multi-scale and multi-orientation features to features, and then a comparison between its features and the
represent and classify faces. An efficient kernel fisher stored features from the training stage is performed to
analysis method was proposed to find multiple subspaces identify the unknown face image. Many algorithms for face
based on both gradient Gabor magnitude and phase features, recognition have both a training stage and a recognition
which is a local kernel mapping method to capture the stage. This phase consists of two main steps that shown in
structure information in faces. Figure 2, including extracting the features for the unknown
Huang, Yi and Pu [17] proposed a new mean shifting image and comparing it with the stored features. The
incremental PCA method based on the autocorrelation calculated projection matrix in the training phase is used to
matrix, which required lower computational time and extract the features of the unknown image. So in order to
storage capacity owing to the two transformation design. extract the features from the unknown image applying PCA
Chougdali, Jedra and Zahid [18] proposed kernel relevance algorithm, multiply the unknown elements of images by the
weighted discriminant analysis for face recognition which projection matrix which is derived from the training stage.
has several interesting characteristics. Two novel kernel When the feature of the unknown image is extracted, a
functions were introduced o improve the performance of the comparison process is performed between the extracted
proposed algorithm. features and the stored features from the training stage.
There are many methods to do such a comparison, such as
3. Face Recognition Algorithms the Euclidean distance measure. The distance between the
features of the unknown image and the stored features is
Face recognition algorithms are implemented in many forms
computed, and then the minimum distance corresponding to
to perform high efficiency of recognition. They are affected
the closest face features is selected as the matched face [17].
by many factors such as recognition rate, execution time and
58 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 8, August 2010
80
75
70
level).
65
Yes Identify
Add the It is
new image matched matching
70
to the the feature vector
database
No
65
1 1.5 2 2.5 3 3.5 4 4.5 5
number of images per person
(c) Recognition rate of 2D - PCA - R before and after applying DWT [4] T. Choudhury, "Three Dimensional Human Face
95
Recognition", Optical Society of India, Vol. 38, No. 1,
pp. 16-21, 2009.
90 .... after applying 2nd level DWT
[5] L. Wang, L. Ding, X. Ding and C. Fang, "2D Face
percentage of the recognition rate
85
-.-. befor applying DWT [9] M. Grgic, K. Delac and S. Grgic, "SCface – Surveillance
Cameras Face Database", Multimedia Tools and
80 Applications, 30 Oct. 2009.
[10] Y. Zhan et al., "Evolutionary Fusion of Multi-Classifier
75 System for Efficient Face Recognition",International
Journal of Control, Automation, and Systems, Vol. 7,
70 No. 1, pp. 33-40, 2009.
[11] A. Saleh et al., "Feature Map Sharing Hypercolumn
65 Model for Shift Invariant Face Recognition", Artificial
1 1.5 2 2.5 3 3.5 4 4.5 5
number of images per person Life Robotics, 5-9 Feb. 2009.
[12] Guoxing et al., "An LPB-Based Multi-Scale
Figure 5. Comparison of recognition rate for different
Illumination Processing Method for Face Recognition",
algorithms.
Journal of Electronics (China), Vol. 26, No. 4, Jul.
2009.
6. Conclusion [13] J. K. Sing, S. Thakur, D. K. Basu, M. Nasipuri and M.
Kundu, "High Speed Face Recognition Using Self
This work examines the four algorithms of face recognition: Adaptive Radial Basis Function Neural Networks",
Principle Component Analysis, Two Dimensional Principle Neural Computing and Applications, 24 Feb. 2009.
Component Analysis in Column direction, Two [14] Kumar et al., "Feature Selection for Face Recognition:
Dimensional Principle Component Analysis in Row A Memetic Algorithm Approach", Journal of Zhejiang
Direction and Two Dimensional Two Directional Principal University Science A, Vol. 10, No. 8, pp. 1140-1152,
Analysis. An algorithm is proposed to reduce the execution 2009.
time of the face recognition algorithms and improve the face [15] Banerjee et al., "Illumination and Noise Tolerant Phase
recognition rate. The execution rate of an algorithm depends Recognition Based on Eigen Phase Correlation Filter
mainly on the number of operations needed to recognize the Modified by Mexican Hat Wavelet", Journal of Optical
indicated face image. Society, Vol. 38, No. 3, pp. 160-168, 2009.
Introducing DWT through PCA algorithms improved the [16] B. Zhangand and Y. Qiao, "Face Recognition Based on
recognition rate up to 5%, also we can conclude that there Gradient Gabor Feature and Efficient Kernel Fisher
are no effective effects on applying the 2nd level DWT. Analysis", Journal of Neural Computing and
Applications, 4 Nov. 2009.
References [17] D. Huang, Z. Yi and X. Pu, "A New Incremental PCA
[1] H. Wang, "Structural Two Dimensional Principal Algorithm with Application to Visual Learning and
Component Analysis for Image Recognition", Machine Recognition", Neural Process Letter, Vol. 30, pp. 171-
Vision and Applications, Jan. 2010. 185, 2009.
[2] E. Norouzi, M. N. Ahmadabadi and B. N. Araabi, [18] K. Chougdali, M Jedra and N. Zahid, "Kernel
"Attention Control with Reinforcement Learning for Relevance Weighted Discriminant Analysis for Face
Face Recognition under Partial Occlusion", Machine Recognition", Pattern Analysis Applications, 9 Apr.
Vision and Applications, Jan. 2010. 2009.
[3] A. Sevcenco and W. Lu, "Perfect Histogram Matching
PCA for Face Recognition", Multidimensional System
and Signal Processing, 14 Jan. 2010.
60 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 8, August 2010
Authors Profile