As shown in the above figure1, the LBP of image are constructed as it gives binary patterns as a result. The Authors [6, 7] are clearly explained the procedure how to get local his- tograms and LBP descriptors.
International Journal of Scientific & Engineering Research, Volume 4, Issue 8, August-2013 1540
ISSN 2229-5518
Mallikarjuna Rao G, Jella Rajesham
Abstract— The complexities associated with Face Recognition continuously attracting the researches to propose new techniques approaches to efficiently address this problem. LBP (Local Binary Patterns) and LAPP (Local Active Pixel Patterns) are some of the local feature extracting approaches for reducing time and space requirements. Authors[6,7] shown that LAPP effectively used Face Recognition on mobile resource constraint environment. In this paper, it is aimed to identify the recognition sensitivity of the local regions through the Weighted Local Active Pixel Pattern(W LAPP) approach. The experimental results using W LAPP on FG-Net Aging Database is confirmed its improved performance compared to the other variants.
Index Terms— Active Pixels, Expression variation, Face Recognition, Illustration conditions, Local Active Pixel Pattern (LAPP), Local
Binary Pattern(LBP), Pose variation, Weight, Weighted Boosting Weighted Local Actvie Pixel Pattern (W LAPP).
—————————— ——————————
ace Recognition is one of the most relevant applications of image analysis. It is a true challenge to build an automated system which equals human ability to recognize faces. Hu-
mans are quite good at identifying known faces, but the skill set is
not sufficient while dealing with a large amount of unknown fac- es. Hence it is warranted to use the machines for dealing these limitations.
However, the strategies[1,2,3,4,5] adopted are not ade- quate for satisfactory recognition in the environments where vary- ing expression, pose and Illumination conditions are inherent to the environment.
Modern Face Recognition algorithms such as Line Edge
Map algorithms[1,2], LBP[3,4,5],LAPP[6,7,8] are proposed to make a grip on lighting factors appeared on the images, edge de- tection is a good approach to shape the facial features that can be hidden due to unusual lighting conditions.
The paper is organized such that basic techniques LBP, LAPP are discussed in sections 2 and 3. Section 4 is concerned about identifying the weights to the local regions and proposal of new approach WLAPP. Experimental discussion is made in sec- tion 5. Conclusions and results are subsequently discussed.
————————————————
• Mallikarjuna Rao G is currently is presently professor in Dept. of Computer Science, Gokaraju Rangaraju Institute of Engineering &Technology(GRIET), Hyderabad, India.
Email: gmr_333@yahoo.com
• Jella Rajesham is currently pursuing Master of Technology in Computer Science &Engineering from Gokaraju Rangaraju Institute of Engineering &Technology, Hyderabad, India, and Pincode: 500090.
Email: rajeshamj@gmail.com
2 Local Binary Patterns (LBP)
The LBP gradually became most widely used approach by researchers and extended it to other pattern recognition disci- plines, which is proposed originally for texture recognition.
As shown in the above figure1, the LBP of image are constructed as it gives binary patterns as a result. The Authors [6, 7] are clearly explained the procedure how to get local his- tograms and LBP descriptors.
Figure 1: LBP Descriptor for 3x3 Mask and Histogram Compu- tation
Though this approach is the one of the efficient ap- proaches for pattern recognition, the researchers are searching for more efficient pattern recognition approaches/Algorithms. This results a new approach called Local Active Pixel Pattern (LAPP), is discussed in next section.
IJSER © 2013 http://www.ijser.org
International Journal of Scientific & Engineering Research, Volume 4, Issue 8, August-2013 1541
ISSN 2229-5518
Authors [6] suggested method for LAPP is capable of support- ing face recognition in both conventional and resource con- straint environment.
The Active Pixel is the one which denotes essential in-
formation of the image and the image can be reconstructed by using only active pixels.
Figure 2: Computation of Active pixels
Computation of Active Pixels uses Brody Transform [6] and extracts the information from images. As a result, an Active pixel matrix is formed. This computation procedure of Active pixels is shown in below Flow chart (figure2).
This process is clearly explained by Authors [6].
Active Template is computed from Active pixel matri-
ces of each person images by taking the average of those Ac- tive pixels of all images belongs to each person (subject). This can be shown below:
ATM = (1/n) ∑APM
Here, ATM-Active Template Matrix
APM-Active Pixel Matrix
Boosting is performed on Active Pixel Matrices and Active Template by taking the maximum valued element in that ma- trix and that added to each non-zero element of that matrix. Then correlation is performed among all the images and all templates. The results obtained by this process are used in the efficient Recognition of images; this can be discussed further in results section.
The proposed approach, is more efficiently working for taken database, is called Weighted Local Active Pixel Pattern, this will discussed in detailed stated with database follows as weight computation process and so on.
Sensitivity of the region on recognition each region is matched (correlated) with associated region and the relationship among weight, sensitivity and Active pixels is shown below:
Figure 3: Weight computing process
The weights of individuals are computed by matching their active pixels with associated template. This process is as fol- lows:
First Active pixel matrices and their template are di- vided into 2x2 matrices.
Then this Active pixel 2x2 matrices are correlated
with corresponding template 2x2 matrices.
IJSER © 2013 http://www.ijser.org
International Journal of Scientific & Engineering Research, Volume 4, Issue 8, August-2013 1542
ISSN 2229-5518
The correlation values can be obtained by predefined MATLAB function r=corr2(A,B), where r is a scalar value.
r=corr2(Activepixels,template)
These correlation values which are >0.3 is considered
and a count is taken for each person individually so that individual count for each region is obtained. This is said to be sensitivity of that region.
Combination of these values for each person is repre- sented by a matrix called weight of the individual person.
This process can be shown in figure3.
A new algorithm called WLAPP is shows the computation of
Weighted Active pixels and Weighted Templates as follows:
First Weight is computed for individuals as explained
in the above section 4.1.
After obtaining weight for each person, Active pixel matrices elements are multiplied by those weights of region as each region contains separate weight.
Here Weighted Active Pixel matrices and Weighted
Template matrices are computed by using Product
Block approach, which means element by element multiplication is performed for matrices instead of matrix multiplication.
This can be called Element-wise multiplication in
MATLAB and represented in MATLAB notation as
follows:
Here, WAPM-Weighted Active Pixel Matrix
APM-Active Pixel Matrix
WATM-Weighted Active Template Matrix
ATM-Active Template Matrix
Now correlation is performed among Weighted Ac- tive Pixel Matrices and Weighted Active Template ma- trices.
This process is called Weighted Boosting.
After that Recognition of images is performed, is discussed in next section.
In traditional approaches, images are to be compressed before the processing to reduce the space requirement. However the compression process results in information loss and influ- enced the recognition accuracy.
In this experimentation to retain the final details of images, enhanced images of size 256x256 are used here so the space requirements are reduced at the same time it gives effec- tive feature set. The local relations are explored using 8 neigh- bourhoods on segmented local blocks whose size is varied. The effect on block size is on recognition accuracy is also in- vestigated.
The FG-NET Aging Database contains face images showing a number of subjects at different ages have been generated as part of the European Union project FG-NET (Face and Gesture Recognition Research Network). The database has been devel- oped in an attempt to assist researchers who investigate the effects of aging on facial appearance.
Figure 4: A person images of FG-Net Aging Database
Face Recognition on FG-Net aging Database is turned into challenging problem as the dataset contains Facial images of the same person taken at different ages (0-65years) with different backgrounds and with different illumination condi- tions of pose variation and varying facial expressions, contains totally 921 images of females and males of different age groups, is shown in below Table1.
Table1: Dataset details
IJSER © 2013 http://www.ijser.org
International Journal of Scientific & Engineering Research, Volume 4, Issue 8, August-2013 1543
ISSN 2229-5518
In this experimentation, all 921 images are considered either for Training or for Testing without any pre-processing and pruning.
The experimentation on different segmentations and
different masks of 256x256 resized images is performed. Those are:
1. 4x4 segmentation and 3x3 mask
2. 8x8 segmentation and 3x3 mask
3. 8x8 segmentation and 5x5 mask
4. 16x16 segmentation and 3x3 mask
5. 16x16 segmentation and 5x5 mask
The results of the above different types are shown in figure 11. Among them, it is found that 8x8 segmentation with
5x5 mask as provided 95.98% accuracy. The following table2 reveals it. (Note that mask is taken as per the Brody transform [6, 7, and 8]).
Table 2: Recognition Accuracy
Methods--> | Without Boosting | Boosting | Weighted Boosting |
Trained im- ages | 921 | 921 | 921 |
Tested im- ages | 921 | 921 | 921 |
Correctly Recognized images | 746 | 696 | 884 |
False Rec- ognized Im- ages | 175 | 225 | 37 |
The accuracy is calculated as correct rate which is obtained from the ratio, number of correctly recognized images to total number of images, is multiplied by 100.
Figure 5: Example Trained, Tested and Recognized images
The figure 5 shows the sample images of 2 sets which are used for recognition and testing. Then recognized images are also shown.
In this paper we are also proposing parallel, scalable model [9] for LAPP. Since the traditional way of programmability and serial execution has reached its limit, the current trend is not only the programmability but also parallelism to meet the growing demands of high performance computing. The presences of multicores in the state of art computing machines enrich their computational power. However, problem solu- tions have to be reorganized to tap the computational power of these machines. As the Visual applications need enormous computational recourses, the research community focuses more on this area and responding with new approach- es/techniques which paralyze traditionally serial methods. Face Recognition is one of them with inherent complexities associated with traditional methods such as PCA, ICA etc., for extracting scalability and parallization. MPI, OMP open source resources support the paralysation of the software. In this re- gard we attempted scalability of LAPP. Three main methods used are:
1. Broadcasting the pixel values among the parallel workers (cores/threads).
2. Performing the pre-processing for LAPP parallel by
all these workers
3. Gathering the computed values to form the feature
vector.
6.1 Parallel Software: Open Source
OpenMP is easier to program and debug than MPI, directives can be added incrementally (called gradual parallelization). It can still run the program as a serial code and serial code statements usually don't need modification. In this code is easier to understand and may be more easily maintained.
OpenMP can only be run in shared memory comput- ers, requires a compiler that supports OpenMP and is mostly used for loop parallelization.
Message Passing Interface (MPI) passes messages to send/receive data between processes, and it is outgrowth of PVM software. MPI runs on either shared or distributed memory architectures that can be used on a wider range of problems than OpenMP. In this each process has its own local variables and distributed memory computers are less expen- sive than large shared memory computers. On the other hand MPI requires more programming changes to go from serial to parallel version, can be harder to debug and performance is limited by the communication network between the nodes.
IJSER © 2013 http://www.ijser.org
International Journal of Scientific & Engineering Research, Volume 4, Issue 8, August-2013 1544
ISSN 2229-5518
The figure 6 shows the work among the worker in
MPI environment.
Figure 6: OpenMP and Message Passing Interface (MPI)
6.2 LAPP: Parallel Approach
The parallel approach for LAPP is based upon its scalability. The image division, neighbourhood extraction and computa- tion are performed in parallel. For an image of size 256x256 is divided into 8x8 sub-regions using 1024 parallel workers. Each sub-region is then processed to compute active pixels using overlapping 3x3 neighbourhood. The parallel approach re- quired 49 workers for each subject. The following figure 7 gives the implementation aspects.
The recognition sensitivity of each sub-region is used while computing weight for the feature value. The weight computa- tion process is illustrated in the section. In this section it is proposed to compute the weight parallel using 128 workers for each subject. Each image contains 16, 8x8 regions and 8 images are considered for each subject. Thus 128 parallel workers (16x8) are used for this computation, while generat- ing the weights of the features. Weighted features for all 82 subjects are obtained similarly. The figure 8 illustrates parallel model for weight computation of a subject.
Figure 7: Local Active Pixel Pattern (LAPP) Feature Extraction
Figure 8: Feature Weight Vector for a subjects
IJSER © 2013 http://www.ijser.org
International Journal of Scientific & Engineering Research, Volume 4, Issue 8, August-2013 1545
ISSN 2229-5518
Weighted Active Pixel Matrix is computed, as discussed in section 3 using 921 parallel workers. Each worker will take 16 feature elements and 16 weights corresponding to subject cov- ering 16 sub-regions. The dot product is computed to form a
4x4 weighted active pixel feature matrix. All 921 images of the database are used while computing the respective feature ma- trix. Figure 9 is showing the process.
The Weighted feature template for each subject is then com- puted using average weighted features of all images of the concerned subject. In this database 8 images are considered for Weighted feature template construction. 656 parallel workers are used for this purpose (82x8).
Figure 9: Weighted Active Pixel Matrix Extraction (WLAPP)
6.5 Parallel Matching for Test Probe
Figure 10 illustrate testing process of the selected image. The Weighted Active Pixel Feature set of the test probe is matched (correlated) in parallel with each of the 82 Weighted Template feature Matrices. The best match is found by selecting highest correlated Weighted Template which in term provides the class to which test probe belongs. The results confirmed its
serial approach by drastically reducing the computational time, shown in Table 3.
Figure10: Test of Images (finding Best match)
The images of each person are correlated with all the persons images in the Dataset are tested in 3 ways including this WLAPP as follows:
The Active pixels of images are taken without any
Pre-processing and correlate with all the images.
Boosting is performed on the Active Pixels and Tem-
plates, and then correlation is performed with all the
images.
Weighted Boosting is performed on the Active Pixels
and Templates, and then correlation is performed with all images.
The Results of above 3 cases are shown in below figure 11.
These results are obtained by 3 different cases are comparatively in increasing order shows the effectiveness of the methods. When comparing the results are obtained by without pre-processing of the images with Boosting are com- paratively good(except some cases) and Weighted Boosting
IJSER © 2013 http://www.ijser.org
International Journal of Scientific & Engineering Research, Volume 4, Issue 8, August-2013 1546
ISSN 2229-5518
posses quite good results and dominated both the cases dis- cussed before.
Weighted Boosting
Boosting
Without boosting
0 50 100
Figure 11: Results comparison
Table 3: Comparison between Serial and Parallel approach
Finally the serial and parallel approaches are compared in the table 3. From this table 3 we can conclude that time is saved as parallel processing is performed and little better recognition rate is found.
We would like to thank management of GRIET for the sup- port extended and director Shri P.S.Raju for constant motiva- tion.
[1] Takács, 1998. Comparing face images using the modified hausdorff distance. Pattern Recogn., 31: 1873-1881
[2] Antitza. D. and J.L. Dugelay, 2011. Frontal-to-side face re- identification based on hair, skin and clothes patches. IEEE International Conference on Advanced Video and Signal-
Based Surveillance, pp: 309-313
[3] Muhammad, S., M. Sajjad, Y.J. Muhammad and
A.A.Muhammad, 2012. Single Image Face Recognition Using
Laplacian of Gaussian and Discrete Cosine Transforms. IAJIT [4] Ran, H., Z. Wei-Shi and H. Bao-Gang, 2011, Maximum cor- rentropy criterion for robust face recognition. IEEE T. Pattern Anal., 33(8): 1561-1575
[5] Joe, Y., W. De Neve, K.N. Plataniotis and Y.M. Ro, 2011.
Collaborative face recognition for improved face annotation in personal photo collections shared on online social networks. IEEE T. Multi., 13(1)
[6] Mallikarjuna Rao G, VijayaKumari G, Babu G.R, Face
Recognition Applications using Active Pixels, International
Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 4, July-August
2012, pp.466-477
[7] Mallikarjuna Rao G, Vijaya Kumari G, Babu G R, Rajesh V,
5023Fast Local Binary Patterns for Efficient Face Recognition/
J. of Sci. and Eng. Vol. 2(2):22-26, Dec. 2011, Mallikarjuna Rao
G. et al. ISSN: 2086
[8] G. Mallikarjuna Rao, Praveen Kumar, G. Vijaya Kumari, Amit Pande, G. R. Babu: An Investigation into the Use of Par- tial Face in the Mobile Environment. ISVC (2) 2011: 526-535
[9] J. Cho, B. Benson, S. Mirzaei, R. Kastner, “Parallelized Ar- chitecture of Multiple Classifiers for Face Detection”, Proc. IEEE International Conference on Application-specific Sys- tems, Architectures, and Processors, 2009, pp. 75-82.
1. Mallikarjuna Rao G, is presently pro- fessor in Dept. of Computer Science, GRIET, Hyderabad. He completed first Masters Degree in Digital Systems from Osmania University in the year 1993
and second masters degree in CSE from JNTU Hyderabad in the year 2002. He is currently pursuing the PhD in Image Pro- cessing area from JNTUH. In his credit there are 1 journal pub- lication and, 6 international conference publications. He won the best Teacher award in 2005 at GNITS. His research inter- ests are Neural Networks, Pattern Recognition.
2. Jella Rajesham, is pursuing Master of Technology in computer Science from GRIET, Hyderabad.
IJSER © 2013 http://www.ijser.org