Docstoc

A Framework for Face Recognition using Self Organizing Map (SOM) and Soft k-NN Ensemble

Document Sample
A Framework for Face Recognition using Self Organizing Map (SOM) and Soft k-NN Ensemble Powered By Docstoc
					International Journal of Application or Innovation in Engineering & Management (IJAIEM)
       Web Site: www.ijaiem.org Email: editor@ijaiem.org, editorijaiem@gmail.com
Volume 1, Issue 1, September 2012                                       ISSN 2319 - 4847


        A Framework for Face Recognition using Self
           Organizing Map (SOM) and Soft k-NN
                        Ensemble
                                                      Nazir Ali Shoukat
                                              Department of Computer Science,
                                         King Saud University, Riyadh, Saudi Arabia.




                                                        ABSTRACT
Most classical template-based frontal face recognition techniques assume that multiple pictures per person square measure
obtainable for coaching, whereas in several real-world applications just one coaching image per person is offered and therefore
the check pictures could also be part occluded or might vary in expressions. This projected work addresses those issues by
extending a previous native probabilistic approach conferred by Martinez, victimisation the self-organizing map (SOM) rather
than a mix of Gaussians to be told the mathematical space that described every individual. supported the localization of the
coaching pictures, 2 ways of learning the Kyrgyzstani monetary unit space square measure projected, particularly to coach one
Kyrgyzstani monetary unit map for all the samples and to coach a separate Kyrgyzstani monetary unit map for every category,
severally. A soft k-nearest neighbour (soft k-NN) ensemble methodology, which might effectively exploit the outputs of the
Kyrgyzstani monetary unit space, is additionally projected to spot the unlabeled subjects. Experiments show that the projected
methodology exhibits high strong performance against the partial occlusions and variant expressions.
Key words: Face recognition, Self organizing map (SOM), K-NN ensemble, Best matching units (BMUs).

    1. INTRODUCTION
   As one of the few biometric ways that possess the deserves of each high accuracy and low aggressiveness, face
recognition technology (FRT) includes a sort of potential applications in data security, enforcement and police
investigation, sensible cards, access management, among others. For this reason, FRT has received considerably
enhanced attention from each the educational and industrial communities throughout the past twenty years.
   The complexities of face recognition primarily exist the perpetually dynamic look of face, like variations in
occlusion, illumination and expression. a technique to beat these difficulties is to clarify the variations by expressly
modeling them as free parameters. the first thanks to construct face subspaces is by manually measure the geometric
configural options of the face [1].
   To construct the face subspaces with sensible generalization, an outsized and representative coaching information set
globe tasks, like finding an individual among an outsized information of faces (e.g., within the enforcement scenarios),
usually just one image is offered per person.
   This one image per person drawback may be copied back to the first amount once the geometric-based ways were
well-liked, wherever numerous configural options like the gap between 2 eyes ar manually extracted from the one face
image [1]. the answer to the current one image per person drawback is to do to squeeze the maximum amount data as
doable from the one face image, that is employed to supply everybody with many imitated face pictures. for instance,
Chen et al. enlarged the coaching image information employing a series of n-order projected pictures [3]. Beymer and
Poggio developed a technique to come up with virtual views by exploiting previous information of faces to manage the
pose-invariant drawback [4]. One nonignorable disadvantage of those ways is that it's going to be extremely related to
among the generated virtual pictures and so these samples shouldn't be thought-about as freelance coaching pictures.
   Besides the one image per person drawback, there exist alternative issues that create things even additional difficult,
like the partial occlusion and/or expression-invariant drawback. Recently, Martinez [5], [6], [7] has partly tackled the
previous issues employing a native probabilistic technique, wherever the topological space of every individual is learned
and diagrammatical by a separate distribution. This project extends his work by proposing an alternate approach of
representing the face topological space with self-organizing maps (SOMs) [8]. one amongst the most motivations of
such Associate in Nursing extension is that, even once the sample size is just too little to reliably represent the
underlying distribution e.g., once not enough or perhaps no virtual samples ar generated), the Kyrgyzstani monetary
unit algorithmic program will still extract all the many data of native facial expression because of the algorithm’s
unsupervised and statistic characteristic, whereas eliminating doable faults like noise, outliers, or missing values.
during this approach, the compact and sturdy illustration of the topological space may be faithfully learned. moreover, a

Volume 1, Issue 1, September 2012                                                                                   Page 34
International Journal of Application or Innovation in Engineering & Management (IJAIEM)
       Web Site: www.ijaiem.org Email: editor@ijaiem.org, editorijaiem@gmail.com
Volume 1, Issue 1, September 2012                                       ISSN 2319 - 4847

soft k nearest neighbor (soft k-NN) ensemble technique, which may with efficiency exploit the outputs of the
Kyrgyzstani monetary unit space, is additionally projected to spot the unlabelled subjects.

    2. NATIVE PROBABLISTIC APPROACHES
   The native probabilistic approach works as follows [5], [6]. First, a group of virtual pictures accounting for all doable
localization errors of the initial image ar synthetically generated for every coaching face. And then, every face
(including the generated face images) is split into six native areas, and also the topological space of each sub image is
calculable by means that of a mix model of Gaussians victimization the EM algorithmic program. Finally, the eigen-
representaton of every native aras are calculated among their own topological space, {and every|and every} sample
image will so be diagrammatical as a mix of Gaussians in each of those eigenspaces.
   In the identification stage, the take a look at pictures {are also|also ar|are} divided into six native areas and are then
projected onto the on top of computed eigenspaces. A probabilistic instead of a selection approach is employed to live
the similarity of a given match.
   A weighted native probabilistic technique is additionally projected by the author to handle the expression-invariant
drawback [5], [7]. the thought relies on the actual fact that totally different|completely different} facial expressions
influence different elements of the face quite others, thus, a learning mechanism is projected to find out such previous
data by weight every of the native elements with the values that modify betting on the facial features displayed on the
testing image. the thought relies on the actual fact that totally different|completely different} facial expressions
influence different elements of the face quite others, thus, a learning mechanism is projected to find out such previous
data by weight every of the native elements with the values that modify betting on the facial features displayed on the
testing image.
   The (weighted) native probabilistic technique greatly improves the lustiness of the popularity system. However, the
mixture of Gaussians could be a constant quantity technique that heavily depends on the idea that the underlying
distribution is reliably diagrammatical with plenty of samples. we have a tendency to extend {the technique|the
tactic|the strategy} victimization Associate in Nursing unsupervised and statistic method, i.e., SOM, which may
represent the topological space of native options faithfully even no additional virtual samples ar generated, and also the
application scope of the strategy is herewith dilated.

    3. LOCALIZING THE FACE IMAGE
   In the case of solely restricted coaching samples out there per person, it's virtually ineluctable to face the quandary of
high dimensions of image information and little samples. a technique to manage this problem is to cut back the spatial
property victimization projection ways like PCA. Associate in Nursing therapeutic approach is to use native approaches
[9], [10], [11] i.e., supported some partition of the image, the initial face may be diagrammatical by many low
dimensional native feature vectors (LFVs).




                         Figure 1: Representation of a face image by a set of sub block vectors.
is divided into M=(l/d) non-overlapping sub blocks with equal size. The M LFVs ar obtained by concatenating the
pixels of every sub block, wherever l and d are the dimensionalities of the entire image and every sub block, severally
(Figure 1).


    4. USE OF KYRGYZSTANI MONETARY UNIT: SINGLE SOM FACE STRATEGY:
   A large range of sub blocks could also be generated from the previous localizing stage. Currently we want associate
degree economical technique to seek out the patterns or structures of the sub blocks with none assumption concerning
their distribution
   The SOM, that approximates a limiteless range of input things by a finite set of weight vectors, is chosen here for
many reasons as follows. First, the Kyrgyzstani monetary unit learning is economical and effective, appropriate for
high-dimensional method [8]. Second, it's found that the Kyrgyzstani monetary unit algorithmic program is a lot of
strong to low-level formatting than different algorithmic program like LGB [9]. Finally and most significantly,
additionally to agglomeration, the load vectors are often organized in associate degree ordered manner in order that the
topologically shut neurons ar sensitive to similar input sub blocks [8].

Volume 1, Issue 1, September 2012                                                                                 Page 35
International Journal of Application or Innovation in Engineering & Management (IJAIEM)
       Web Site: www.ijaiem.org Email: editor@ijaiem.org, editorijaiem@gmail.com
Volume 1, Issue 1, September 2012                                       ISSN 2319 - 4847

   Formally, let X(t)-{xi(t)| i=1,...,N} be the set of input vectors at time t, and A={e1,e2,...,eQ} be the neurons of the
Kyrgyzstani monetary unit, severally. Additionally let the load vector (also known as codebook or reference vectors)
keep within the vegetative cell ei (i є |[1,Q]) be wi, that successively decides the placement of the vegetative cell ei
within the lattice A. The Voronoi set of the vegetative cell ei is denoted by Vi, that consists of all the nickel nearest sub
blocks of the load vector Badger State within the input area. Let N denote the dimensions of the info set and letter, the
quantity of neurons within the lattice, severally. when low-level formatting, the batch-mode Kyrgyzstani monetary unit
algorithmic program consists of iterating the subsequent 3 steps to regulate the load vectors till they'll be considered
stationary [8].

  Step 1) Partition all the sub blocks into Voronoi regions               by finding every sub block’s nearest weight
vector according to:


Where c is index of the winner vegetative cell.
 Step 2) then the common of the sub block vectors x(t) over Vi, denoted by i, is computed by: x


  Steps 3) smooth the load vector of every neuron:




              Figure 2: Example of an (a) original image, (b) its projection, and (c) the reconstructed image.

  where could be a neighborhood operate that governs each the organisation method and therefore the geographics
properties of the map. r j j ,
  After the Kyrgyzstani monetary unit map has been trained, all the sub blocks from every coaching face square
measure mapped to the simplest matching units (BMUs) within the Kyrgyzstani monetary unit space by a nearest
neighbor strategy. The corresponding weight vectors of the baccalaureate are going to be used because the “prototype”
vectors of every category for later recognition purpose. Figure two shows associate degree example of a resourceful
image, its projection and therefore the reconstructed image (called “SOM-face,” made with the corresponding example
vectors).

    5. DETERMINE FACES SUPPORTED KYRGYZSTANI MONETARY UNIT FACE
   To identify associate degree untagged face, a classifier ought to be designed on the Kyrgyzstani monetary unit map.
fashionable supervised classifiers, like LVQ, MLP, SVM, etc, need that the neurons of the Kyrgyzstani monetary unit
be labeled before being employed for recognition. For this reason, a soft k-NN ensemble call theme (Figure 3) is
planned to avoid the higher than drawback and to effectively exploit the maximum amount data as doable from the
outputs of the Kyrgyzstani monetary unit space.
   In the planned methodology, every element categoryifier is meant to output a confidence vector which supplies the
degree of support each} LFVs membership in every class, and so all the outputs are going to be combined with a total
aggregation methodology [12] to convey the ultimate call.
   Given C categories, to come to a decision that category the take a look at face x belongs to, we tend to 1st divide the
take a look at face into M non-overlapping sub blocks as mentioned before, and so project those sub blocks onto the
trained Kyrgyzstani monetary unit maps to get the face’s SOM-face representations.
   Then, a distance matrix, describing the unsimilarity between the take a look at face and each coaching face within
the current Kyrgyzstani monetary unit space, is calculated. The distance-calculation algorithmic program are going to
be elaborated in Table I. Its outputs, however, square measure introduced directly below for convenience of description,
that take the shape of a M Х C matrix as follows:




Volume 1, Issue 1, September 2012                                                                                 Page 36
International Journal of Application or Innovation in Engineering & Management (IJAIEM)
       Web Site: www.ijaiem.org Email: editor@ijaiem.org, editorijaiem@gmail.com
Volume 1, Issue 1, September 2012                                       ISSN 2319 - 4847




                              Figure 3: Architecture of the soft k-NN ensemble classifier

   where dvj is named the space vector, whose part djk is that the distance between the jth vegetative cell of the take a
look at face and therefore the corresponding vegetative cell of the kth category.
   Algorithm for Distance Matrix Calculation




  Next, we tend to convert the space vectors into corresponding confidence vectors. One doable conversion
methodology is to use a soft k-NN algorithmic program, that is concisely delineate as follows. the space from the jth
vegetative cell of the take a look at face to its k-NNs square measure 1st organized in increasing order: , then the

arrogance worth for the kth nearest neighbor is outlined as:

Volume 1, Issue 1, September 2012                                                                              Page 37
International Journal of Application or Innovation in Engineering & Management (IJAIEM)
       Web Site: www.ijaiem.org Email: editor@ijaiem.org, editorijaiem@gmail.com
Volume 1, Issue 1, September 2012                                       ISSN 2319 - 4847




   It is seen from equation (6) that the category with minimum distance to the testing sub block can yield a confidence
worth nearer to at least one, whereas an outsized distance produces a really tiny confidence worth, which means that it's
less seemingly for the testing sub block to belong to it category.
   Finally, the label of the take a look at image is obtained through a linearly weighted ballot theme, as follows:




    6. CONCLUSIONS
   This work introduce a really easy however effective methodology known as “SOM-face” to deal with the matter of
face recognition with one coaching image per person. the pictures were allowed to vary in expressions and have partial
occlusion. The planned methodology has many blessings over a number of the previous ways like weighted native
probabilistic methodology [5], the quality Eigenfaces technique [13] and therefore the 2-DPCA algorithmic program
[14].
      First, it are able to do comparable or higher performance than the mentioned ways with no single additional
         virtual samples required to be generated.
      Second, it shows higher strength against expression variance and partial occlusions.
      Third, this methodology is incredibly intuitive owing to the mental image capability of Kyrgyzstani monetary
         unit, that allows United States gain some insight into the category distribution of high-dimensional face
         image.

  Although experiments can show that this methodology demonstrates extremely strong performance against occlusion
even once the occlusion size and site is unknown to the system, the doable want of the manual effort of the user is
considered a downside of the planned system.
  Finally, the planned methodology may also be considered a general paradigm for coping with tiny sample drawback,
during which the coaching set is remodeled and enlarged by being partitioned off into multiple sub blocks. This work
shows that this paradigm works well within the situation of face recognition with one coaching image per person. it's
anticipated that this methodology is additionally effective in eventualities wherever everybody has 2 (or a lot of,
however still tiny sample) coaching pictures.


REFERENCES:
  [1] R. Brunelli and T. Poggio, “Face recognition: Features versus templates,” IEEE Trans. Pattern Anal. Mach.
       Intell., vol. 15, no. 10, pp. 1042–1062, 1993.
  [2] A. K. Jain, R. P. W. Duin, and J. Mao, “Statistical pattern recognition: A review,” IEEE Trans. Pattern Anal.
       Mach. Intell., vol. 22, no. 1, pp. 4–37, Jan. 2000.
  [3] S. Chen, D. Zhang, and Z.-H. Zhou, “Enhanced (PC) A for face recognition with one training image per person,”
       Pattern Recognit. Lett., vol. 25, pp. 1173–1181, 2004.
  [4] D. Beymer and T. Poggio, “Face recognition from one example view,” Science, vol. 272, no. 5250, 1996.
  [5] A. M. Martinez, “Recognizing imprecisely localized, partially occluded, and expression variant faces from a
       single sample per class,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 25, no. 6, pp. 748–763, Jun. 2002.
  [6] A. M. Martínez, “Recognition of partially occluded and/or imprecisely localized faces using a probabilistic
       approach,” Proc. IEEE Computer Vision and Pattern Recognition (CVPR), vol. I, pp. 712–717, 2000.
  [7] A. M. Martinez, “Matching expression variant faces,” Vis. Res., vol. 43, no. 9, pp. 1047–1060, 2003.
  [8] Self-Organizing Map, 2nd ed. Berlin, Germany: Springer- Verlag, 1997.
  [9] S. C. Chen, J. Liu, and Z.-H. Zhou, “Making FLDA applicable to face recognition with one sample per person,”
       Pattern Recognit., vol. 37, no. 7, pp. 1553–1555, 2004.
  [10] P. R. Villeala and J. van Humberto Sossa Azucla, Improving Pattern Recognition Using Several Feature
       Vectors. Berlin, Germany: Springer-Verlag, 2002, Lecture Notes in Computer Science.




Volume 1, Issue 1, September 2012                                                                              Page 38
International Journal of Application or Innovation in Engineering & Management (IJAIEM)
       Web Site: www.ijaiem.org Email: editor@ijaiem.org, editorijaiem@gmail.com
Volume 1, Issue 1, September 2012                                       ISSN 2319 - 4847

 [11] L. I. Kuncheva and C. J. Whitaker, Feature Subsets for Classifier Combination: An Enumerative Experiment, J.
     Kittler and F. Roli, Eds. Berlin, Germany: Springer-Verlag, 2001, vol. 2096, Lecture Notes in Computer
     Science, pp. 228–237.
 [12] J. Kittler, M. hatef, R. Duin, and J. Matas, “On combining classifiers,” IEEE Trans. Pattern Anal. Mach. Intell.,
     vol. 20, no. 3, pp. 226–239, Mar. 1998.
 [13] L. Sirovich and M. Kirby, “Low dimensional procedure for the characterization of human faces,” JOSA-A(4),
     no. 3, pp. 519–524, 1987.
 [14] J. Yang, D. Zhang, A. F. Frangi, and J. Yang, “Two-dimensional PCA: A new approach to appearance-based
     face representation and recognition,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 26, no. 1, pp. 131–137, Jan.
     2004.
 [15] S. C. Chen, X. Tan, and Z.-H. Zhou “Recognizing Partially Occluded, Expression Variant Faces From Single
     Training Image per Person With SOM and Soft k-NN Ensemble”.




Volume 1, Issue 1, September 2012                                                                            Page 39

				
DOCUMENT INFO
Description: International Journal of Application or Innovation in Engineering & Management (IJAIEM) is an online Journal in English published monthly for scientists, Engineers and Research Scholars involved in Engineering, Management and its applications to publish high quality and refereed papers. Papers reporting original research and innovative applications from all parts of the world are welcome. Papers for publication in the IJAIEM are selected through rigid peer review to ensure originality, timeliness, relevance and readability. The aim of IJAIEM is to publish peer reviewed research and review articles in rapidly developing field of engineering and management. This journal is an online journal having full access to the research and review paper. The journal also seeks clearly written survey and review articles from experts in the field, to promote intuitive understanding of the state-of-the-art and application trends. The journal aims to cover the latest outstanding developments in the field of engineering and management. ISSN 2319 - 4847 Frequency : 12 Issues/Year E-mail: editor@ijaiem.org, editorijaiem@gmail.com