Multimodal Approach for Face Recognition using 3D-2D Face Feature Fusion

Naveen S & Dr. R.S Moni
International Journal of Image Processing (IJIP), Volume (8) : Issue (3) : 2014 73
Multimodal Approach for Face Recognition using 3D-2D Face
Feature Fusion
Naveen S nsnair11176@gmail.com
Assistant Professor, Dept. of ECE
LBS Institute of Technology for Women
Trivandrum, 695012, Kerala, India
Dr R.S Moni moni2006rs@gmail.com
Professor, Dept. of ECE
Marian Engineering College,
Trivandrum, Kerala, India
Abstract
3D Face recognition has been an area of interest among researchers for the past few decades
especially in pattern recognition. The main advantage of 3D Face recognition is the availability of
geometrical information of the face structure which is more or less unique for a subject. This
paper focuses on the problems of person identification using 3D Face data. Use of unregistered
3D Face data for feature extraction significantly increases the operational speed of the system
with huge database enrollment. In this work, unregistered 3D Face data is fed to a classifier in
multiple spectral representations of the same data. Discrete Fourier Transform (DFT) and
Discrete Cosine Transform (DCT) are used for the spectral representations. The face recognition
accuracy obtained when the feature extractors are used individually is evaluated. The use of
depth information alone in different spectral representation was not sufficient to increase the
recognition rate. So a fusion of texture and depth information of face is proposed. Fusion of the
matching scores proves that the recognition accuracy can be improved significantly by fusion of
scores of multiple representations. FRAV3D database is used for testing the algorithm.
Keywords: Point Cloud, Rotation Invariance, Pose Correction, Depth Map, Spectral
Transformations, Texture Map and Principal Component Analysis.
1. INTRODUCTION
3D Face recognition has been an active area of research in the past decades. The complications
encountered in the enrollment phase and the huge computational requirements in the
implementation phase have been the major hindrance in this area of research. The scenario has
improved tremendously due to the latest innovations in 3D imaging devices and has made 3D
Face recognition system a reliable option in security systems based on Biometrics. Though poor
resolution is a major drawback encountered in 3D Face images the geometrical information
present in 3D facial database can be exploited to overcome the challenges in 2D face recognition
systems like pose variations, bad illumination, ageing etc.
In this work, focus is made on an identification problem based on 3D Face data using fusion
schemes. Identification corresponds to the person recognition without the user providing any
information other than the 3D facial scan. The system arrives at an identity from among the
enrolled faces in the database. Use of texture information along with the geometrical information
of the face seems to improve the recognition accuracy of face recognition system when pose
correction is not done as a preprocessing step.
Alexander M. Bronstein, Michael M. Bronstein and Ron Kimmel [2] proposed an idea of face
recognition using geometric invariants using Geodesic distances. C. Beumier [3] utilized parallel

planar cuts of the facial surfaces for comparison. Gang Pan, Shi Han, Zhaohui Wu and Yueming
Wang [4] extracted ROI of facial surface by considering bilateral symmetry of facial plane. Xue
Yuan, Jianming Lu and Takashi Yahagi [5] proposed a face recognition system using PCA, Fuzzy
clustering and Parallel Neural networks. Trina Russ et al [6] proposed a method in which
correspondence of facial points is obtained by registering a 3D Face to a scaled generic 3D
reference face. Ajmal Mian, Mohammed Bennamoun and Robyn Owens [7] used Spherical Face
Representation for identification. Ondrej Smirg, Jan Mikulka, Marcos Faundez-Zanuy, Marco
Grassi and Jiri Mekyska [8] used DCT for gender classification since the DCT best describes the
features after de-correlation. Hua Gao, Hazım Kemal Ekenel, and Rainer Stiefelhagen[9] used
Active Appearance model for fitting faces with pose variations. Mohammad Naser, Moghaddasi
Yashar Taghizadegan and Hassan Ghassemian [10], used 2D-PCA for getting the feature matrix
vectors and used Euclidean distance for classification. , Omid Gervei, Ahmad Ayatollahi, and
Navid Gervei[11] proposed an approach for 3D Face recognition based on extracting principal
components of range images by utilizing modified PCA methods namely 2D-PCA and
bidirectional 2D-PCA. Wang et al.,[14] described point signatures in a 3D domain and facial
feature points by using Gabor filter responses in a 2D image. Chang et al., [15] tested the
recognition algorithm using fusion of 3D and 2D information and was found effective in improving
face recognition rate(FRR). C. McCool et al., [16] used Log-Gabor Templates of depth and
texture data was used along with Mahalanobis Cosine metric as the distance measure and in [17]
PCA difference vectors are modeled using Gaussian Mixture Models (GMMs) and is compared
with Mahalanobis Cosine metric. Pamplona Segundo et al., [18] used boosted cascade classifiers
using range images as input for scale invariant face detection for the automation of face
recognition systems. Jahanbin S et al., [19] 2-D and 3-D Gabor coefficients and the
anthropometric distances are calculated and three parallel classifiers are fused at the match
score level to form a face recognition system.
A typical 3D Face is shown in Figure.1. Figure.2 represents its axis level representation. Figure 3
and 4 represents the texture map with two different orientations.
FIGURE 1: 3D Face Model. FIGURE 2: 3D Face in Space.
FIGURE 3: 2D Texture Map (Frontal View). FIGURE 4: 2D Texture Map (Left Turned Head).

First the fusion of the representative transformations from 3D to 1D space and 3D to 2D space is
considered. Since only a sparse set of points in the 3D point cloud are available, it is necessary to
increase the data density by using multiple data representations generated from same raw data.
For this the data is transformed into spectral domain using DFT and DCT. This sparse set of data
with occlusion can be effectively countered by invoking multiple score fusion schemes which can
effectively improve the feature data density. Use of Depth information alone is not sufficient for an
efficient recognition system since pose correction is not done. So texture information is also
incorporated with the fusion scheme.
2. PROPOSED IDEA FOR FACE RECOGNITION
FIGURE 5: Proposed Method .
The system aims at extracting the feature from the input data through multiple feature extraction
tools and fuses the scores to get a system with better recognition accuracy. The main feature
extraction principle used in this system is the spectral transformation. The spectral transformation
tools used here are 1D-DFT and 2D-DFT along with 2D-DCT. These spectral transformations
transform the data to a better representation which increases the accuracy of recognition system.
The most important part of this work lies in the pattern classification problem. A pattern of data
points is available. This pattern is not sufficient for the recognition system to work since the data
will be highly occluded due to pose variations in the X, Y and Z axis or in any complex plane.
The 3D Face recognition scheme is affected by pose variations of the subject (person under
consideration). There are methods available in which the correction to this effect of pose
variations also is included. One such method is the Iterative Closest Point (ICP) algorithm. But
the main disadvantage of these methods is that a reference face is to be used as a model for
other rotated faces to be corrected. Also the processing time taken is very high. Further, the
reliability of this result depends on the accuracy in selection of the reference face model used.
Therefore, in this work done, this correction to the effect of pose variation is not considered. The
method aims at recognizing the subject without much computationally complex mathematical
procedure. Also the results prove that the efficiency of system is comparable with a system with
pose correction. The idea behind spectral representation of data is that, when data is in spatial

domain, comparison will be done as one to one pixel level or voxel level. So the rotation and
translation of data will highly affect the result. Moreover the accuracy of the system will go down
to even 5% under severe pose variations in X, Y and Z axis. When spectral transformation is
done the distributed data will be concentrated or it may be represented in a more uniform way.
I.e. the input data will be concentrated and represented uniformly in spectral domain. The
translation and rotation invariance properties of the transformations used will aid to improve the
accuracy of system significantly.
Here FRAV3D database is considered. It contains the facial data with different face orientations
and expressions. When depth information alone was considered the Face recognition accuracy
(FRA) was not high. So texture data of face is also considered which significantly improves the
FRA.
The data available for the analysis and testing will be in Point Cloud format which is a matrix
array of size Mx3. The value of M denotes the number of points in the 3D space. For each data
input the number of points used will be different for representation. An optimum number of points
are selected for Point Cloud data vector. Texture information is available as RGB image of size
400x400. This image is converted to intensity image and is down sampled by a factor 2 to reduce
the computation time. Down sampling doesn’t reduce the FRA, it is verified.
The proposed method involves the following steps given below in sequence. As a preliminary
step Z component of the point cloud data alone is taken and DFT is applied on this data to get
spectral representation. Second step is to map the 3D points to 2D grid without pose correction to
get the depth map. From this 2D depth map nose tip is detected using Maximum Intensity Method
and the area around the nose (ROI-Region of Interest) is extracted (Figure 7 and Figure 8). On
this ROI data 2D-DFT and 2D-DCT is applied. Simultaneously spectral representation of Texture
map is also taken using 2D-DCT and 2D-DFT. Once spectral representations are obtained,
Principal Component Analysis (PCA) is applied on that data to get the corresponding weight
vectors. This weight vectors are fed to a classifier which uses Euclidean distance for
classification. Here 2D PCA is used for depth and texture features and 1D PCA for Z component
of point Cloud.
2.1 Point Cloud Representation as a One Dimensional Vector
We have the point cloud data as an M x 3 matrix. Of this the third dimension, which is the depth
information alone is taken. This reduces the number of points under consideration to M while the
original being 3M. This also aids the real time implementation of the system faster. We call this
data as Metric.
The Metric data will be distributed in spatial domain as shown in Figure 6. It relate to the data of
Z axis with some unknown orientation and scale. The data distribution is different for different axis
at different orientation. But when DFT is computed, it turns out that the data distribution becomes
similar. This significantly will improve the recognition accuracy. The reason for choosing DFT over
DCT which has a better energy compaction property is that with DCT, the recognition accuracy
with X axis and Y axis rotation has been found to be almost half of that obtained using DFT.
Transformation of Metric data to spectral domain can be done using 1D-DFT, using equation (2).
F(K)= 



1
0
/2
)(
N
n
NKnj
enf 
, for a vector of size Nx1 (2)

FIGURE 6: Metric in Spatial Domain.
2.2 Nose Tip Localization and Face Area Extraction
For localizing the nose tip, maximum intensity method is used. In this method assumption is
made that the nose tip will be the point with maximum pixel intensity. Once the nose tip is found
the circular area (ROI) around the nose tip is extracted using an optimum radius. Now the depth
map will contain the face area only, all other unwanted portions are cropped away. Next face area
is centralized by making the nose tip as the center pixel of the image. Otherwise the matching
process will result in a lower accuracy. The face area is also normalized by the maximum
intensity. The centralized face image is as shown in Figure 7 and Figure 8. Figure 9 represents
the depth map obtained from point cloud data oriented in different axis.
FIGURE 7: Depth Map. FIGURE 8: ROI from Depth Map.
FIGURE 9: Depth Map with different pose variations along X, Y and Z axis.
2.3 Use of 2D DFT and 2D DCT on Depth Data and Texture Map
The point cloud data in 3D space is projected to an X-Y grid to get 2.5D (2.5D image is the depth
map itself) image using standard projection formula. This depth data will be having the pixel value
as the Z- Coordinate of Point Cloud data. Face images have higher redundancy and pixel level
correlation which is a major hindrance in face recognition systems. Transforming face images to

spectral domain will reduce the redundancy. Here only the magnitude of spectral data is taken
alone since it is not transformed back to spatial domain in any of the processing stages.
Now, the Depth image is transformed to spectral domain using 2D-DCT. The energy compaction
will take place and the result will be again an M x N matrix. The result is shown in Figure 10. Here
the global feature extraction by DCT is made used of. DCT has the property of de-correlation
which enables the data structure to loose spatial pixel dependency. The low frequency
components which mainly form the facial features will be prominent in the transformed space
which makes the pattern classification more reliable, since the human eyes are more sensitive to
information in low frequency spectrum.
Transformation to spectral domain using 2D-DCT can be done using equation (3) and the
transformed image is shown in Figure.10.
F(u,v)= )
2
)12(
cos()
2
)12(
cos(),(
1
0
1
0 N
vy
M
ux
yxf
N
y
M
x
 





, for a M x N depth image. (3)
Again the depth image is transformed using 2D-DFT so that rotation effects are reduced. DFT is a
rotation invariant transformation. So that the distributed pixel values (normalized) are properly
aligned, this enables the pattern matching more efficient. DFT spectrum of face image will appear
as shown in Figure 11. Transformation to spectral domain using 2D Discrete Fourier
Transformation can be done using equation (4).
F(U,V)= 





1
0
)//(2
1
0
),(
N
y
NVyMUxj
M
x
eyxf 
, for a M x N depth image (4)
Now the error score is estimated using all the multiple representations separately. For processing
metric, 1D-PCA is used and for processing 2D DCT and 2D DFT representation 2D-PCA is used.
1D-PCA was also checked for the 2D representations but 2D-PCA gave better result for 2D
representations.
FIGURE 10: DCT Representation of ROI Depth Map. FIGURE 11: DFT Representation of ROI Depth Map.
The same procedure is repeated for the texture map also to get the spectral representation of
gray scale intensity image using 2D-DCT as shown in Figure 12 and using 2D-DFT as in
Figure 13.

FIGURE 12: DFT Representation of Texture Map. FIGURE 13: DCT Representation of Texture Map.
2.4 Principal Component Analysis
Use of spectral transformations will make the data samples almost uncorrelated. Even then,
some spatial dependency may exist. So Principal Component Analysis (PCA)[12] is used for the
correlation process as it uses orthogonal transformations to get linear uncorrelated data sets
called principal components. Conventional covariance method for the calculation of principal
components is used here. Feature extraction using 1D-PCA is done as follows.
Let Xi be the spectral transformed 1D Euclidean Metric which representations i
th
person, it is
grouped as a M x N matrix X=[X1 X2 …XN], where N is the number of face samples under
consideration.
Mean vector is calculated as follows
Xm= 
N
i
iX
N 1
1
(5)
Standard deviation will be calculated
XSD= 

N
i
mi XX
N 1
)(
1
(6)
Covariance matrix is calculated
XCOV= XSD* X
T
SD (7)
This is a matrix of size M x M, which is of very large dimension. Also it gives M Eigen values and
M Eigen vectors which are very large in number to process. The base idea of dimensional
reduction by changing the construction of covariance matrix can be now used.
XCOV= X
T
SD* XSD (8)
The result is a matrix of size N x N, where N is the number of subjects under consideration. It
gives N Eigen values and N Eigen vectors. The Eigen values are sorted in descending order and
will select the first N’ largest Eigen values and corresponding Eigen vectors. Eigen vectors in N’
dimension is transformed to the higher dimension of M by multiplying with Standard deviation
Matrix. Now the test data is projected to this lower dimension space to get the corresponding
weight vectors.

In 2D-PCA 2D spectral representation of Depth map is considered. The only difference in
calculating the Covariance matrix is that here a 2D matrix is used when compared to 1D Matrix in
1D-PCA. After determining the Eigen values and Eigen vectors a 2D weight vector matrix is
obtained which is then converted to a column matrix.
2.5 Score Fusion
Error values are calculated for each data representations and all this error values are combined
as a single error value using the linear expression as given in equation9.
W= [W_DMDCT W_DMDFT W_PCDFT W_TMDCT W_TMDFT] (9)
E= [Error_DMDCT Error_DMDFT Error_PCDFT Error_TMDCT Error_TMDFT] (10)
Error= W*E
T
(11)
On the equation 11 W_DMDCT is weight for error value obtained using 2D DCT on depth data
and is its value is taken as 1 , W_DMDFT and W_PCDFT are weights for the error value obtained
using 2D DFT and 1D DFT on depth map and metric respectively. This weight values are
selected in such way that the error values of 2D DFT and 1D Euclidean metric are in same scale
as that of error value due to 2D DCT. Similarly texture score weights W_TMDCT, W_TMDFT are
the weights of error score obtained using 2D DCT and 2D-DFT on intensity image. Weight value
can be approximated using equation 12, 13, 14 and 15.
W_DMDFT
1
=
Error_DCT
Error_DFT
, rounded to 10
2
’s. (12)
W_PCDFT
1
=
Error_DCT
TError_PCDF
, rounded to 10
9
’s. (13)
W_TMDCT
1
=
Error_DCT
TError_TMDC
≈1 (14)
W_TMDFT
1
=
Error_DCT
TError_TMDF
, rounded to 10
2
’s (15)
These values give optimum recognition accuracy. Finding an optimum weight for this error
function which can minimize it can be another optimization problem.
3. RESULTS
3.1 Results of Individual Representation Accuracy Analysis
For analysis and testing FRAV3D database is used here. It contains 106 subjects. Of these 100
subjects were taken into consideration. Testing was done on the input data with pose and
orientations as Frontal, Right turn 25º (respect to Y axis), Left turn 5º (respect to Y axis), Severe
right turn (respect to Z axis), Soft left turn (respect to Z axis), Smiling face, Open mouth, Looking
upwards (turn respect to X axis), Looking downwards (turn respect to X axis), Frontal view with
lighting changes etc. A substantial change in the recognition accuracy is observed with fusion
scheme.
First recognition accuracy obtained with Depth data and Point cloud metric will be considered.
Accuracy is checked by varying the number of faces from 10 to 100, with enrollment increment of

2 faces and a summary is shown in Table 1. When the DCT of Depth Map is analyzed the
accuracy variance is about 9.12% with maximum of 54.44% and minimum of 45%, for DFT it is
14.40% with maximum of 57.29 % and minimum of 46.67% and for Point Cloud metric it is
17.13 % with maximum of 60 % and minimum of 43.33%.
No of
Subjects
Total
Face
Samples
Tested
Depth
Map DCT
Depth
Map DFT
Metric
DFT
Accuracy Accuracy Accuracy
60 960 48.33 48.33 43.33
90 1440 54.44 56.67 45.56
100 1600 53.00 56.00 46.00
TABLE 1: Comparison of Accuracy obtained using Depth Map Spectral Transformation and Point Cloud
Metric.
When the reliability of a Face recognition system is considered, the system should correctly
classify the users based on the current pose and posture of the user. A user can be enrolled in
the data base of the FR system mostly in limited number of pose/ expression variations. As the
number of enrollment sessions increases the system will be trained in such a way that the user
can be correctly identified in a latter testing session. But it does have a limitation unless the
training sessions given are sufficient which covers all the pose, translation and expression
variations.
Here such an analysis is done to check the reliability of the system. The summary results are
shown in Table 2.
When Table 2 is analyzed, it can be observed that all 16 sessions are not identified by any of the
feature extraction scheme. For example, 7 sessions are successfully identified only by 39
enrolled faces when DCT is used, 46 when DFT is used and 43 when point cloud metric is used
out of the total 100 enrolled faces. Obviously this brings down the overall recognition accuracy of
the system and hence the reliability. In other words only 273, 322 and 301 samples out of 1600
samples were successfully identified while others are rejected.

Depth
Map DCT
Depth
Map DFT
Point
Cloud
DFT
Sessions Subjects Subjects Subjects
1 100 100 100
2 96 96 96
3 87 91 85
4 82 81 69
5 69 73 55
6 53 61 49
7 39 46 43
8 28 30 33
9 19 22 24
10 14 13 14
11 7 3 10
12 4 1 5
13 2 1 1
14 0 1 1
15 0 0 0
16 0 0 0
TABLE 2: Comparison of number of enrolled faces successfully identified in 16 different sessions.
3.2 Results of Fusion Scheme Analysis
As explained earlier, when Texture Map Score obtained by the spectral transformation of RGB
data using 2D-DCT and 2D-DFT is also fused with the score obtained using depth data notable
improvement in recognition accuracy is observed. Those observations are shown in Table 3. It
can be also noted that the accuracy deviation with final fusion scheme is nearly 4.17%, i.e. the
system is reliable even when more number of faces are enrolled and face recognition accuracy
won’t drop too much.
No of
Subjects
Total
Face
Samples
Tested
Fusion
DM
FUSION
of DM
and TM
Accuracy Accuracy
60 960 50.31 74.58
90 1440 52.36 76.88
100 1600 52.44 78
TABLE 3: Comparison of FRR of Texture fusion scheme with FRR of Depth Fusion scheme.
When Table 3 is analyzed the accuracy variance for Depth Score Fusion is also around 4.06%. In
short fusion of individual score of different representations reduces the accuracy variance and
hence increases the reliability of the FR systems.
Table 4 shows the difference in number of subjects identified with fusion of Texture and Depth
spectral representation scores and depth alone.

Fusion
DM
Fusion of
DM and
TM
Sessions Subjects Subjects
1 100 100
2 100 100
3 99 100
4 97 100
5 92 100
6 83 100
7 80 95
8 67 87
9 61 82
10 50 73
11 41 66
12 30 55
13 20 37
14 13 25
15 5 14
16 1 3
TABLE 4: Comparison of number of enrolled faces successfully identified in each sessions.
All testing are done only with a single image enrolled in data base as the reference Image. When
the images in the database itself are fed as the test image the accuracy is 100% itself. All other
samples are tested as the class that is not present in the database, which ensures that the
accuracy obtained in the proposed system is insensitive to the session variations and least
sensitive to pose and expression variations. The insensitivity is very clear from the accuracy
obtained using fusion schemes. Even when the number of samples is doubled, say from 800 to
1600, the FRR change is near to 2% only. Testing with 160 samples which contains complex
pose variations and expressions gave 80% accuracy and 1600 samples gave an accuracy of
78% which ensure that the system is reliable.
The main reason for the reduction in accuracy is due to the loss of laser signals while scanning
the face and the occlusions due to pose variations [13]. The advantage of proposed work is that
face alignment or face registration is not done here. The registration process takes sufficient
computation time and it is not acceptable in real time systems. This system requires only a single
image as training image for achieving this much accuracy. By increasing the number of training
images FRR can be further improved. This method seems to perform better when compared to
[10] where the recognition rate is only 70% with a single training image. While 2D PCA [11]
method gives 76.2% when tested over 540 samples.
3.3 Improvement In Accuracy Rates by Fusion Scheme
When the accuracy obtained in the fusion scheme is analyzed, when depth data is fused with
texture information a significant improvement is observed as shown in Table 5.

No of Subjects
Correct samples
additionally
recognized
Fusion DM and
TM
Improvement in
FRR
60 464 48.23
70 507 45.21
90 675 46.83
100 769 48.04
TABLE 5: Improvement in FRR with fusion of Depth and Texture information.
Almost an average improvement of nearly 48% can be obtained when the fusion scheme is used.
3.4 Face Recognition Rate-Graphical Comparison
FIGURE 14: FRR Comparison of Individual and Fusion Scheme.
3.5 Computation Time
Testing of algorithm is done on 3GHz, Core I-5 processor; the average identification time per
sample is 400ms.This will again increase as the number of subjects increase. By using down
sampling and abstracting the data representations computational time can be further reduced.
But with a dedicated system, the testing time can be further reduced to microseconds.
4. CONCLUSIONS
The fusion algorithm is tested on unregistered 3D Faces with different pose orientations and
expressions. The algorithm gives an accuracy of 78% even with unregistered face data.
Experiments are conducted on various representation types and feature extraction methods for
3D Face recognition. The experimental results show that the features can be effectively extracted
from depth data, point cloud and texture map using spectral transformation like DFT and DCT.
Fusion experiments were conducted at score level. The surface representation of 3D Face data in
terms of metric with its spectral transformation, in addition to the spectral representation of
projected depth information to the XY plane and Intensity image as texture Map using 2D-DFT
and 2D-DCT are also used. When the numbers of subjects in the database are increased, the
recognition rate remained more or less same which shows that the further enrollment of faces in

the database won’t affect the true recognition rate. There is ample scope for further improvement
using more fusion schemes at the representation level and at spectral level. This method can be
implemented in real time systems since the processing time required is lesser on a dedicated
system. Dimensional reduction method can also improve the time performance of the system.
Advancement of technology in 3D Face capturing and faster processing systems can make the
system more efficient in all aspects.
Acknowledgment
Thanks are due to Dr. Enrique Cabello, Universidad Rey Juan Carlos, Spain, for providing us with
the FRAV3D database.
5. REFERENCES
[1] http://www.frav.es/databases/FRAV3D
[2] Alexander M. Bronstein; Michael M. Bronstein and Ron Kimmel. “Expression-invariant 3D
face recognition.” In Proc. International Conference on Audio- and Video-based Biometric
Person Authentication, volume 2688 of Lecture Notes in Computer Science Guildford, UK,
2003, pp:62-70.
[3] C. Beumier, “3D face recognition” In IEEE Int. Conf. on Computational Intelligence for
Homeland Security and Personal Safety (CIHSPS2004), Venice, Italy, Jul 2004, pp:21-22.
[4] Gang Pan; Shi Han Zhaohui Wu and Yueming Wang. “3D Face Recognition using Mapped
Depth Images.” Proceedings of the IEEEComputer Society Conference on CVPR (CVPR’05)
Workshops- Volume- 03, 2005 p:175.
[5] Xue Yuan;Jianming Lu and Takashi Yahagi. “A Method of 3D Face Recognition Based on
Principal Component Analysis Algorithm.” IEEE International Symposium on Circuits and
Systems, Vol. 4 May 2005. pp: 3211 - 3214.
[6] Trina Russ; Chris Boehnen and Tanya Peters. “3D Face Recognition Using 3D Alignment for
PCA”, Proceedings of the 2006 IEEE Computer Society Conference on CVPR (CVPR’06)
Volume 2, 2006 pp: 1391 – 1398.
[7] Ajmal Mian; Mohammed Bennamoun and Robyn Owens. ”Automatic 3D Face Detection,
Normalization and Recognition.” Proceedings of the Third International Symposium on
3DPVT (3DPVT'06) Jun 2006, pp: 735-742.
[8] Ondrej Smirg, Jan Mikulka, Marcos Faundez-Zanuy, Marco Grassi and Jiri Mekyska. “Gender
Recognition Using PCA and DCT of Face Images.” Advances in Computational Intelligence ,
Lecture Notes in Computer Science Volume 6692, 2011, pp: 220-227.
[9] Hua Gao, Hazım Kemal Ekenel and Rainer Stiefelhagen. “Pose Normalization for Local
Appearance-Based Face Recognition.” Advances in Biometrics, Lecture Notes in Computer
Science Volume 5558, 2009, pp:32-41.
[10]Mohammad Naser-Moghaddasi Yashar Taghizadegan and Hassan Ghassemian.(2012, Feb),
.”3D Face Recognition Method Using 2DPCA-Euclidean Distance Classification”, ACEEE
International Journal on Control System and Instrumentation(Vol 3), Available:
http://hal.archives-ouvertes.fr/docs/00/74/16/40/PDF/70.pdf.
[11]Omid Gervei, Ahmad Ayatollahi and Navid Gervei.”3D Face Recognition Using Modified PCA
Methods” World Academy of Science, Engineering & Technology; Mar 2010, Issue 39, p264
[12]M. Turk and A. Pentland, “Eigenfaces for recognition”, J. Cognitive Neuroscience , 1991,
3(1), pp. 71 – 86.
[13]Image Analysis and Recognition, Third International Conference, ICIAR 2006, Póvoa de
Varzim, Portugal, September 18-20, 2006, Proceedings, Part II, Lecture Notes in Computer
Science , Volume 4142 2006.

[14]Y. Wang, C. Chua and Y. Ho, “Facial feature detection and face recognition from 2D and 3D
images”, Pattern Recognition Letters, vol. 23, 2002, pp. 1191-1202.
[15]K. I. Chang, K. W. Bowyer and P. J. Flynn, “Multiple nose region matching for 3D face
recognition under varying facial expression”, IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 28, no. 10, 2006, pp. 1695-1700.
[16]C. McCool, J. Cook, V. Chandran and S. Sridharan, “Combined 2D / 3D Face Recognition
using Log-Gabor Templates”, Proceedings of the IEEE International Conference on Video
and Signal Based Surveillance (AVSS'06), 2006, pp. 83.
[17]C. McCool, J. Cook, V. Chandran and S. Sridharan, “Feature Modelling of PCA Difference
Vectors for 2D and 3D Face Recognition”, in Proceedings of IEEE International Conference
on Video and Signal Based Surveillance, 2006, pp. 57.
[18]Pamplona Segundo, M , Silva, L and Bellon, O.R.P, “ Real-time scale-invariant face detection
on range images” , in IEEE International Conference on Systems, Man and Cybernetics
(SMC), 2011 , pp:914 - 919.
[19]Jahanbin, S, Hyohoon Cho, Bovik, A.C, “Passive Multimodal 2-D+3-D Face Recognition
Using Gabor Features and Landmark Distances”, in IEEE Transactions on Information
Forensics and Security(Volume:6,Issue: 4 ), 2011, pp: 1287 - 1304.

Multimodal Approach for Face Recognition using 3D-2D Face Feature Fusion

Recommended

More Related Content

What's hot (18)

Viewers also liked (7)

Similar to Multimodal Approach for Face Recognition using 3D-2D Face Feature Fusion (20)

Recently uploaded (20)

Multimodal Approach for Face Recognition using 3D-2D Face Feature Fusion