Automatic Analysis of Smoothing Techniques by Simulation Model Based Real-Time System for Processing 3D Human Faces

International Journal of Embedded Systems and Applications (IJESA) Vol.4,No.4, December 2014
DOI : 10.5121/ijesa.2014.4402 13
AUTOMATIC ANALYSIS OF SMOOTHING
TECHNIQUES BY SIMULATION MODEL BASED
REAL-TIME SYSTEM FOR PROCESSING 3D HUMAN
FACES
Suranjan Ganguly1
, Debotosh Bhattacharjee2
and Mita Nasipuri3
Department of Computer Science and Engineering, Jadavpur University, India
ABSTRACT
The pivotal research work that has been carried out and described in this literature acknowledges the
importance of various smoothing techniques for processing 3D human faces from 2.5D range face images.
The smoothing techniques have been developed and implemented using MATLAB-Simulink for real time
processing in embedded system. In addition, the significance of smoothed 2.5D range image over original
face range image has been discovered as well as its time complexity has also been reported with array of
experiments. The variations in time complexities are also accomplished using different optimization levels
and execution modes. A set of filtering techniques such as, Max filter, Min filter, Median filter, Mean filter,
Mid-point filter and Gaussian filter, have been designed and illustrated using Simulink model. The model
takes depth face image (i.e. the range face image) as input in real time and presents the improvement over
original face images. In the design flow, the performance of every block has also been characterized by
range face images from Frav3D, GavabDB, and Bosphorus databases. In the experimental section of this
research article, an array of performance analysis for these smoothing techniques with variation of
frameworks is explained.
KEYWORDS
3D face image, 2.5D face image, MATLAB-Simulink, Smoothing techniques, Range face image
1. INTRODUCTION
Computer vision based different methodologies like object recognition, registration,
identification, etc. deploys the 2D or 3D face images into automation system. Hence, the growth
of image scope, and variation of applications require the computation of a complex image
processing methodologies. But, sometimes these algorithms lack behind due to the presence of
noise, outliers, spikes, holes, etc. For this reason, some important image data is suppressed, or
lost, or some noisy data get itself processed that leads to poor performance of the particular
mechanisms.
The images may incorporate variations of noises due to acquisition problems, quantization or
digitization error or scanning error, etc. Now, it is very much required to filter out these noises
and smooth the facial surface of the input face images for practical use of the algorithm in real
time applications. In this context, the development and implementation of different linear as well
as non-linear filtering techniques [1] namely: Max filter, Min filter, Mid-point filter, Mean filter,

International Journal of Embedded Systems and Applications(IJESA) Vol.4,No.4,December 2014
14
Gaussian filter, and Median filter have been applied on 3D human face images and it will be
advantageous for further processing.
Human face images are considered to be more reliable biometric feature for automatic security
system for crucial properties such as uniqueness, universality, well accepted and well
understandable by people. It is always visible, and every one must have a face whereas other
biometric features like hand geometry, ear, eye may be lost due to some reason. The surveillance
cameras are also used to capture the human faces. Hence, face recognition [2-4] got most of the
researcher’s attention from last two decades.
In addition, there is a vast influence to prefer 3D human face [4] images rather than 2D images.
Specifically, the 2D images preserve the reflectance characteristics of the object in the pixel data.
So, it is mainly dependent on the illumination variations whereas 3D face images are particularly
used to preserve the depth values in X-Y plane. Another property that makes 3D face images
more convenient than 2D is 3D geometrical rotation along X, Y, and Z axes. Thus, the pose
variation, the major problem of current face recognition, can be resolved using face registration
[5-6] mechanism.
However, the states of the art of filtering techniques in case of 3D face processing reason have
been summarized in table 1. In this literature study, its importance in face registration and (or)
recognition has particularly an impact for developing an array of smoothing techniques
implementing in real time system and illustrating their significance for processing purpose.
Table 1. The state-of-the-art of image smoothing techniques for 3D face images.
Reference Description
[7] Authors have demonstrated the effect of the median filter for removing sharp
spikes, and again interpolation technique has been added to fill the holes on the
face image.
[8] Authors have compared the performance of landmark localization technique with
array of smoothing methods, namely Max Filter, Min filter, Gaussian filter, Mean
filter, and Weighted median filter.
[9] Here, authors have used median and Gaussian filter for smoothing purpose. The
median filter is used for spikes from 3D faces and again, Gaussian filtering is
applied for removing surface noise.
[10] To detect the nose-tip, authors have computed Gradient Weighting Filter method
during the smoothing process of their proposed algorithm.
2. MOTIVATION AND APPLICATION
Studying the recent state of the art regarding the influence of smoothing techniques for 3D human
face processing, authors have proposed an approach to real time processing of some of the
filtering techniques using MATLAB-Simulink model.
2.1. Range Image creation
The 2.5D range [11] face images are gray like face images. The difference between gray 2D and
2.5D is that, 2.5D images are comprised by depth values (or Z’s values) from 3D images where as
2D images are intensity values. Thus, the background has minimum depth value i.e. zero (0) and
nose region (especially ‘pronasal’) landmark has a maximum depth [6] [12] value 255. In figure

15
1, 2D, 2.5D and 3D face images of randomly selected subjects from Frav3D database [13] is
described.
2D Face image 2.5D Face image 3D Face image
Figure 1. 2D, 2.5D and 3D face images of Frav3D database
Other than Frav3D face database, GavabDB [14] and Bosphorus [15] face databases have also
been considered for emphasizing the significance of smoothing technique using a simulation
model [19-20] of the embedded system. In figure 2, created range face images of randomly
selected subject from GavabDB and Bosphorus database have been illustrated in figure 2.
(a) From GavabDB database (b) From Bosphorus database
Figure 2. Created range face image
2.2. Smoothing algorithms
During the investigation phase, authors have implemented spatial linear as well as order-statistic
[1] [16] (i.e. non linear) filters on depth values of 2.5D range face images. The linear filters [17]
specifically an Mean filter and Gaussian filter are computed whereas in order statistic
categorization of image filtering, Median filter, Max filter, Min filter, and Midpoint filter are
applied on range face images.
2.2.1. Preprocessing technique
Before, these series of filters are experimented on depth values for their significance, a
preprocessing task have been carried out. The range images have been padded by zeros in the
opposite side of each row and column of the image. Thus, each and every depth values from the
furthest row and column of the image can be processed for better performance analysis.
Otherwise, it would not be considered during spatial image processing purpose. This phenomenon
is shown in figure 3. In this figure, a block of depth values with 8×8 grid from a section of 2.5D
range face image is shown.

16
(a) without padding (b) with zero padding
Figure 3. The importance of depth image padding
The highlighted (circled by yellow color) depth value is processed first by the smoothing
techniques. If the smoothing technique is used on original range image then, the far most two
rows and columns will be unchanged, whereas with padding zeros with the original image will
effectively affect these sections. This padding is done in real time and again it has been removed
after the filtering techniques have been applied. Thus, the dimension is preserved before and after
smoothing technique.
.
2.2.2. Smoothing by linear filter
Linear filters do not depend on any kind of order of depth values (or intensity) from filtering
kernel. The filters in this category only compute the linear functions (like Gaussian or Averaging)
for removing the noises irrespective of the ordering of values encompassed by the filtering
window.
Gaussian filter: It is an important filter among set of smoothing filters from linear class. The
weight of the Gaussian filter [1] [16] is chosen from the Gaussian kernel. For the qualitative
measurement during this research work for 2.5D depth face image, 2D Gaussian kernel is
implemented. The kernel function [16] with ߪ = 3 is computed. It is observed that, a large value
of ߪ i.e. variance has the wider filter and smoothing impact.
Mean filter: Mean filter [1] is simple linear spatial filter that averages the neighbor’s depth values
of the filter mask. It is also referred as low pass filter [18]. To analyze the effect of the averaging
or mean filter for depth face images, a 3×3 kernel have been undertaken.
2.2.3. Smoothing by nonlinear filter
These filters are also known as order-statistic filter [1]. It is a nonlinear smoothing filter whose
output is emphasized on the ordering of the values encompassed by the filtering mask. Now, the
output from the ranking result is used to modify the center depth value of the mask. Here, for
nonlinear order statistic filter, authors have also considered for computed 3×3 kernel filter mask.
Max filter: In this noise filtering mechanism, 100% or highest depth value from neighborhood
depth values is chosen. Hence, for depth based image filtering scheme, the holes (containing
minimum or ‘0’ depth value) may be removed.
Min filter: It is useful to select the minimum or 0% depth value among the selected data by the
filtering window. Hence, the spikes (containing maximum depth) within the human face surface
due to scanning error can be minimized.
Mid-point filter: It is another type of smoothing technique which is used to select the depth value
in between maximum and minimum. It is a similar type of Mean filter as described above.

17
Median filter: It is one of the famous and well known order-statistic filtering scheme where, 50%
among 9-values selected by 3×3 filtering window. It has another qualitative property that, it
provides us less blurring effect [1] than linear filters.
The different outputs from these filters have been demonstrated in the discussion section where
the significance of each output is broadly discussed.
2.3. Discussion
In this section, the outcomes of respective filters of randomly selected subjects from three
databases are shown in figure 4.
Frav3D
database
GavabDB
database
Bosphorus
database
After Smoothing Significance
Gaussian Smoothing
Mean filter
Max filter
Min filter
Mid-point filter
Median filter
Figure 4. Visualization of smoothing effects

18
In this observation, it is noticed that the smoothing technique has a great significance over depth
values of the human face. So, its application in real time may significantly improve different
aspects of processing human face like face registration, recognition, etc.
Now, from Gaussian analysis it is noted that the outer portion is blurred out more than other. It
actually shrinks to center as it is dense and simultaneously blurs the edges. Hence, depth values
near different facial regions like eyes, eyebrows, the nose region, and lips all have this quality. It
is the property of the Gaussian filter, and it has successfully been executed in real time for depth
values. In the case of another linear filter namely, Mean filter the edges are preserved. In these
points, the depth value is nearly same as the average value that has been computed by 3×3
window. In might have a greater significance for landmark localization, face component
extraction, etc. The same significance has also been found for Mid-point filtering technique. By
mathematical logic it is determining 50% i.e. in between maximum and minimum, likely same as
average filter. For Max and Min filter, authors have observed the same significance after
smoothing technique. The reason of such significant output is that it is either selecting 0% or
100% depth value under the filtering window. Thus, it almost has a binary thresholded image as
shown in [11]. Hence, the spikes and holes can be removed in this process. The most well known
order statistic Median filtering method preserves the elliptical concave and convex curve details
near eye region, the nose region, lip region, etc.
3. MODEL DESIGN AND IMPLEMENTATION
Model has been designed and implemented using MATLAB-Simulink environment. Different
modules from Simulation tool have been coupled to finalize the implemented model. The
detailing of the blocks has been illustrated later in this section. It is an approach for real time
human computer interaction for visualizing the significant effect of different filtering techniques
on 3D human faces. Not only model design, successful code-generation has also been attempted.
In figure 5, developed model has been described.
Figure 5. Illustration of developed models
MAX /MIN / MEAN/
GAUSSIAN/MEDIAN/
MID-POINT Filter

19
In the figure, the zoomed region shows different user defined embedded MATLAB function
blocks that have been designed for real time processing in embedded system.
In general, three ‘User-Defined MATLAB Function Block’, four ‘Sinks’ blocks, one ‘Constant
Source’, two ‘Math Operations’ block are incorporated with the advanced model. Each of these
blocks has its own crucial role for successful implementation of an advanced model.
The 3D face images [11] are far more different from 2D images as described earlier in the
motivation section. Instead of intensity values, depth (values along Z-axis) in X-Y plane is
preserved in 3D images in ‘.wrl’, ‘.abs,' ‘.bnt’ like formats. Hence, before the smoothing
technique is applied on 2.5D range face image, it has been generated from 3D face image using
an ‘Interpreted MATLAB Function’ (shown in upper-left corner of figure 5). The ‘Constant
Source’ is used as input of depth values to the Simulation model. Now, ‘MATLAB Function
Blocks’ are allowed to embed the source code for displaying the range face image and then it is
processed, and significance has been highlighted. The ‘Sinks’ have been used to produce an
output from each block for better human computer interaction. Finally, two mathematical
operations are used for real time 2D matrix manipulation purpose.
After, the successful implementation of ‘Simulation’ model, it is further required to generate code
for embedded system. For this purpose successful code generation have been accomplished by
choosing ‘C-language’ as target language with ‘Optimization on’ parameter of Compiler
optimization level, ‘Fixed-step’ solver option (i.e. fixed step size of 0.02) along with ‘Auto
generated comments’. In figure 6, code generation report for Mid-point filter is shown.
Figure 6. Description of the code generation report
In this figure, the expression of ‘ZI’, the constant block from the model, is highlighted. It contains
the depth values from range image which have been used during the execution time of the model.

20
4. EXPERIMENTAL RESULT
Though there are other parameters for validating the model for real time application, but authors
have used two variations among them. One is Simulation mode i.e. whether the model is tested by
Normal mode or Accelerator or Rapid Accelerator mode and another parameter is Compiler
Optimization level (either Optimizations off or Optimizations on).
In table 2, an array of analysis of this execution parameters with simulation stop time 1.0 is
summarized. This description is used for the range image that has been selected randomly from
Frav3D database. It has been tested on 4-GB RAM with Windows 7 (64-bit) professional
Operating System environment and Inrel i5-3470 CPU with 3.20GHz processors.
[
Table 2. Performance analysis of different parameter configuration
Smoothing Techniques Simulation Mode
Simulation
Stop time 1.0
Compiler Optimization Level
Optimizations off Optimizations on
Gaussian Filter
Normal 13.967028 seconds 9.988573 seconds
Accelerator
9.907918 seconds
[After Successfully built
the Accelerator target]
9.813664 seconds
[After Successfully
built the Accelerator
target]
Rapid Accelerator
12.664001 seconds
the rapid accelerator
target]
11.924632 seconds
[After Successfully
built the rapid
accelerator target]
Mean filter
Accelerator
6.954723 seconds
6.503646 seconds
[After Successfully
target]
Rapid Accelerator
9.511578 seconds
target]
9.285953 seconds
[After Successfully
built the rapid
accelerator target]
Max filter
Accelerator
6.528666 seconds
6.706566 seconds
[After Successfully
target]
Rapid Accelerator
10.056241 seconds
target]
9.704065 seconds
[After Successfully
built the rapid
accelerator target]
Min filter
Accelerator
6.622307 seconds
6.610166 seconds
[After Successfully
target]
Rapid Accelerator
9.467616 seconds
target]
10.232370 seconds
[After Successfully
built the rapid
accelerator target]
7.109918 seconds 7.186364 seconds

21
Mid-point filter
Accelerator [After Successfully built
[After Successfully
target]
Rapid Accelerator
10.006641 seconds
target]
9.296432 seconds
[After Successfully
built the rapid
accelerator target]
Median filter
Accelerator
12.267989 seconds
12.691126 seconds
[After Successfully
target]
Rapid Accelerator
10.047549 seconds
target]
9.809040 seconds
[After Successfully
built the rapid
accelerator target]
From this outline, it is noticed that the complexity is much higher for the techniques which are
required mathematical computation much higher than others. The Mid-point filter takes more
time than Max and (or) Min filter, whereas Median filter also accounts more time for smoothing
operation in real time. It requires ordering of depth values encompassed by the filtering window.
The Gaussian filter also consumes more time to process 3D human face image for real time
application.
In figure 7, a comparative study is shown among the time complexities of different smoothing
methods with an array of parameters arrangement.
Figure 7. Comparison of the performance study

22
Selecting the two restrictions such as Simulation mode and Compiler optimization level is
explained here. The ‘Normal’, ‘Accelerator’ and ‘Rapid Accelerator’ modes are defined to
compare the time complexities side by side for better understanding of the minimum time span of
the specified filtering technique. Along with this, with the Compiler optimization parameter set-
up, fastest execution (by Optimizations on) and fastest compilation (by Optimizations off) has
also been represented.
5. CONCLUSIONS
The applications of smoothing techniques during image processing or computer vision (especially
in face recognition) have a crucial implication. In this literature, authors have explained the
influence of some of the smoothing techniques 2.5D face images. In addition, authors have
created a real time application model by MATLAB-Simulink model and validated by series of
parameters’ composition. The code generation has also been conducted and implemented. Along
with this, the validation of the model is done on three modern 3D face databases (namely Frav3D,
GavabDB, and Bosphorus) having two different 3D image formats like ‘.wrl’ and ‘.bnt’.
Now, authors are focusing to implement these methods for range face images using Field
Propagation Gateway Array (FPGA) to develop better dedicated system with much lesser time
complexity.
ACKNOWLEDGEMENTS
Authors are thankful to a project supported by DeitY (Letter No.: 12(12)/2012-ESD), MCIT,
Govt. of India, at Department of Computer Science and Engineering, Jadavpur University, India
for providing the necessary infrastructure for this work.
REFERENCES
[1] Gonzalez, R. C., And Woods, R.E., (2007) “Digital Image Processing,” 3rd Edition, Prentice Hall
Publisher.
[2] “Face Recognition,” Pp. 1-10, August 2006, Url: ttp://Www.Biometrics.Gov/Documents/Facerec.Pdf.
[3] Jelsovka, D., Hudec, R., Breznan, M., Kamencay, P., (2012) “2d-3d Face Recognition Using Shapes
Of Facial Curves Based On Modified Cca Method”, 22nd International Conference Radioelektronika
(Radioelektronika), Pp. 1-4.
[4] Ganguly, S., Bhattacharjee, D., And Nasipuri, M., (2014) “3d Face Recognition From Range Images
Based On Curvature Analysis”, Volume: 04, Issue: 03, Pp. 748-753.
[5] Ayyagari, V.R., Boughorbel, F., Koschan, A., Abidi, M.A., (2005) “A New Method For Automatic 3d
Face Registration”, Proceedings Of The Ieee Computer Society Conference On Computer Vision And
Pattern Recognition (Cvpr’05), Pp. 1-8.
[6] Ganguly, S., Bhattacharjee, D., And Nasipuri, M., (2014) “Range Face Image Registration Using Efri
From 3d Images”, In Advances In Intelligent And Soft Computing, Springer, Accepted In
Proceedings Of 3rd Frontiers Of Intelligent Computing: Theory And Applications (Ficta 2014).
[7] Soltana, W. B., Ardabilian, M. Lemaire, P., Huang, D., Szeptycki, P., Chen, L., Erdogmus, N.,
Daniel, L., Dugelay, J., Amor, B.B., Drira, H., Daoudi, M., Colineau, J., (2012) “3d Face
Recognition: A Robust Multi-Matcher Approach To Data Degradations”, In Proc Of Icb 2012, Pp
103-110.
[8] Bagchi, P., Bhattacharjee, D., Nasipuri, M., & Basu, D.K. (2012) “A Novel Approach To Nose-Tip
And Eye-Corners Detection Using H-K Curvature Analysis In Case Of 3d Images”, In Proc Of
International Journal Of Computational Intelligence And Informatics, Vol. 2: No. 1.
[9] Hatem, H., Beiji, Z., Majeed, R., Lutf, M., Waleed, J., (2013) “Nose Tip Localization In Three-
Dimensional Facial Mesh Data”, International Journal Of Advancements In Computing Technology
(Ijact), Volume5, Number13,Pp. 99-105.

23
[10] Margret N. Silva, Vipul Dalal, (2013) “ Nose Tip Detection Using Gradient Weighting Filter
Smoothing,” International Journal Of Engineering Research And Development, Volume 9, Issue 5,
Pp. 09-11.
[11] Ganguly, S., Bhattacharjee, D., And Nasipuri, M., (2014) “2.5d Face Images: Acquisition, Processing
And Application”, Computer Networks And Security, International Conference On Communication
And Computing (Icc-2014), Organized By Alpha College Of Engineering, India, Publisher: Elsevier
Science And Technology, Pp. 36-44 Isbn: 978935107244.
[12] Dhane, P., Jain, A., Kutty, K. K., (2011) “A New Algorithm For 3d Object Representation And Its
Application For Human Face Verification”, International Conference On Image Information
Processing, Pp. 1-6.
[13] Frav3d Face Database, Url: Http://Www.Frav.Es/Databases/Frav3d/
[14] Gavabdb Face Database, Url: Http://Gavab.Escet.Urjc.Es/Recursos_En.Html
[15] Bosphorus Face Database, Url: Http://Bosphorus.Ee.Boun.Edu.Tr/Default.Aspx
[16] Jayaraman, S., Esakkirajan,S., And Veerakumar, T., (2010), “Digital Image Processing”, 3rd Edition,
Tmh Publisher.
[17] Linear Filters, Url: ttp://Luthuli.Cs.Uiuc.Edu/~Daf/Courses/Cs5432009/Week%203/Simplefilters.Pdf
[18] Spatial Filters - Mean Filter, Url: Http://Homepages.Inf.Ed.Ac.Uk/Rbf/Hipr2/Mean.Htm
[19] Ganguly, S., Bhattacharjee, D., And Nasipuri, M., (2014) “Analyzing The Performance Of Haar-
Wavelet Transform On Thermal Facial Image Using Matlab-Simulink Model”, Proceedings Of The
1st International Conference On Microelectronics, Circuit And Systems,: Volume 2, Pages: 106-111,
Isbn:81-85824-46-0.
[20] Tfrs Using Simulink, Url: Https://Www.Youtube.Com/Watch?V=3l-Qd2zv5xs&Feature=Youtu.Be
AUTHORS
SURANJAN GANGULY received the M.Tech (Computer Technology) degree from
Jadavpur University, India, in 2014. He completed B-Tech (Information Technology) in
2011. His research interest includes image processing, pattern recognition. He was a
project fellow of UGC, Govt. of India, sponsored major research project at Jadavpur
University. Currently, he is a project fellow of DietY (Govt. of India, MCIT) funded
research project at Jadavpur University.
BHATTACHARJEE received the MCSE and Ph.D. (Eng.) degrees from Jadavpur
University, India, in 1997 and 2004 respectively. He was associated with different
institutes in various capacities until March 2007. After that he joined his Alma Mater,
Jadavpur University. His research interests pertain to the applications of computational
intelligence techniques like Fuzzy logic, Artificial Neural Network, Genetic Algorithm,
Rough Set Theory, etc. in Face Recognition, OCR, and Information Security. He is a life
member of Indian Society for Technical Education (ISTE, New Delhi), Indian Unit for
Pattern Recognition and Artificial Intelligence (IUPRAI), and a senior member of IEEE (USA).
MITA NASIPURI received her B.E.Tel.E., M.E.Tel.E., and Ph.D. (Engg.) degrees
from Jadavpur University, in 1979, 1981 and 1990, respectively. Prof. Nasipuri has been
a faculty member of J.U since 1987. Her current research interest includes image
processing, pattern recognition, and multimedia systems. She is a senior member of the
IEEE, U.S.A., Fellow of I.E. (India) and W.B.A.S.T, Kolkata, India.

Automatic Analysis of Smoothing Techniques by Simulation Model Based Real-Time System for Processing 3D Human Faces

Recommended

More Related Content

Similar to Automatic Analysis of Smoothing Techniques by Simulation Model Based Real-Time System for Processing 3D Human Faces (20)

More from ijesajournal (20)

Recently uploaded (20)

Automatic Analysis of Smoothing Techniques by Simulation Model Based Real-Time System for Processing 3D Human Faces