Eye detection using wavelets and ANN

EYE DETECTION USING WAVELETS AND ANN

Project with matlab code

 ABSTRACT

 A  Biometric  system  provides  perfect  identification  of  individual  based  on  a  unique biological  feature  or  characteristic  possessed  by  a  person  such  as  finger  print,  hand

writing,  heart  beat,  face  recognition  and  eye  detection.  Among  them  eye  detection  is  a better  approach since Human  Eye  does not  chan ge throu ghout the life  of  an  individual.  It is regarded as the most reliable and accurate biometric identification system available. 

         In  our  project  we  are  going  to  develop  a  system  for  ‘eye  detection  using  wavelets and  ANN’  with  software  simulation  package  such  as  matlab  7.0  tool  box  in  order  to verify  the  uniqueness  of  the  human  eyes  and  its  performance  as  a  biometric.  Eye

detection  involves  first  extracting  the  eye  from  a  digital  face  image,  and  then  encoding the  unique  patterns  of  the  eye  in  such  a  way  that  they  can  be  compared  with  pre-

registered  eye  patterns.  The  eye  detection  system  consists  of  an  automatic  segmentation system  that is  based on  the wavelet  transform,  and  then  the Wavelet analysis  is used  as  a pre-processor  for  a  back  propagation  neural  network  with  conjugate  gradient  learning.

The  inputs  to  the  neural  network  are  the  wavelet  maxima  neighborhood  coefficients  of face images  at  a  particular scale.  The  output  of the neural  network  is the classification of the input into an eye or  non-eye  region.   An accuracy of  81% is  observed f or test images under different environment conditions not included during training.

 

Eye  detection  system  is  being  extensively  used  in  biometrics  security  solutions  by  U.S. Department  of  Defense  (DOD),  which  includes  access  control  to  ph ysical  facilities,

security  systems  or  information  databases.  Suspect  tracking,  surveillance  and  intrusion detection  and  by  various  Intelligen ce  agencies  through  out  the  world,  also  in  the

corrections/laws enforcement marketplaces. 

1.1. Biometric Technology: 

A  biometric  system  provides  automatic  recognition  of  an  individual  based  on  some  sort of  unique  feature  or  characteristic  possessed  by  the  individual.  Biometric  systems  have been developed based on fingerprints, facial features,  voice, hand  geometry, handwriting,

the retina, and the one presented in this project, the eye. 

Biometric  systems  work  by  first  capturing  a  sample  of  the  feature,  such  as  recording  a digital  sound  sign al  for  voice  recognition,  or  taking  a  digital  color  image  for  eye

detection.  The  sample is  then transformed  using some sort  of  mathematical  function into a  biometric  template.  The  biometric  template  will  provide  a  normalized,  efficient  and highly  discriminating  representation  of  the  feature,  which  can  then  be  objectively

compared  with  other  templates  in  order  to  determine  identity.  Most  biometric  systems allow two  modes  of operation. A  training  mode  or enrolment  mode  for  adding  templates

to  a  database,  and  an  identification  mode,  where  a  template  is  created  for  an  individual and then a match is searched for in the database of pre-en rolled templates. 

A  good  biometric  is  characterized  by  use  of  a  feature  that  is;  highly  unique  so  that  the chance  of  any two  p eople  having  the  same  characteristic  will  be  minimal,  stable  so  that the  feature  does  not  change  over  time,  and  be  easily  captured  in  order  to  provide

convenience to the user, and prevent misrepresentation of the feature. 

1.2. EYE: The Perfect ID

The  randomness  and  uniqueness  of  human  eye  p atterns  is  a  major  breakthrough  in  the search  for  quicker,  easier  and  highly  reliable  forms  of  automatic  human  identification, where the human eye serves as a type of 'biological passport, PIN or password’. 

Results  of  a  study  by John  Daugman  and Cathryn,  of over  two million  different  pairs  of human eyes in images taken from  volunteers in Britain, USA and  Japan show  that no two

eye patterns  were  the  same  in  ev en  as much  as  one-third of  their  form.  Even genetically

identical  faces  -  for  example from  twins  or  in  the  probable  future,  from  human  clones  - have different eye patterns.

The implications of eye detection are highly signif icant at  a time when organizations such as  banks  and  airlines  are  looking  for  more  effective  security  measures.  The  possible

applications  of eye detection  span  all aspects of  daily  life,  from  computer  login, national border  controls  and secur e access to  bank cash machine accounts, to  ticket-less  air  travel,

access  to  premises  su ch  as  the  home  and  office,  benefits  entitlement  and  credit  card authentication. 

Compared with other  biometric technologies,  such as face, speech and finger recognition, eye recognition can easily be considered as the most reliable form of biometric. However,

there  have been  no  independent  trials  of  the  technology,  and  source  cod e  for  systems  is not available in working condition. 

 

1.3. Objective: 

                             The  objective  will  be  to  implement  an  open-source  eye  detection system  in  order  to  verify  the  claimed  performance  of  the  technolog y.  This  project  is

based  on  a  novel  method, which  is robust  and efficient in  extracting  eye  windows  using Wavelets  and  Neur al  Networks.  Wavelet  analysis  is  used  as  a  pre  processor  for  a  back propagation  neur al  network  with  conjugate  gradient  learning.  The  inputs  to  the  neural

network  are  the  wavelet  maxima  neighborhood  coefficients of face  images at a particular scale.  The  output  of  the  neural  network  is  the  classification  of  th e  input  into  an  eye  or non-eye region. The  updated weight  and bias values  for a  particular  person is  stored  in  a

database.  The  image  to  be  verified  is  wavelet  transformed  before  being  applied  to  the neural network with  those  updated weight and  bias  values.  The  person is  identified when

the  neural  network  output  of  one  of   the  test  images  matches  with  that  of  the  verified image.  .    An  accuracy  of  90%  is  observed  for  test  images  under  different  environment conditions  not  included  during  training.

 2.1.    INTRODUCTION

 The  transform  of  a  signal  is  just  another  form  of  representing  the signal.  It  does not  change  the  information  content  present  in  the  signal.  The  Wavelet

transform provides  a  time-frequency rep resentation  of the signal. It was developed  to  overcome the shortcoming  of  the  Short  Time  Fourier  Transform  (STFT),  which  can  also  be  used  to analyze non-stationary signals.  While STFT gives a constant resolution  at  all frequencies,

the  Wavelet  Transform  uses  multi-resolution  technique  by  which  different  frequenciesare analyzed with different resolutions. 

A  wave  is  an  oscillating  function  of   time  or  space  and  is  periodic.  In  contrast, wavelets  are  localized  waves.  They  have  their  ener gy  concentrated  in  time  or  space  and are suited to analysis  of transient  signals.  While  Fourier  Transform  and  STFT use  waves to analyze signals, the Wavelet Transform uses wavelets of finite energy. 

The  wavelet  analysis  is  done  similar  to  the  STFT  an alysis.  The  signal

to be analyzed  is  multiplied  with  a  wavelet  function  just  as  it  is  multiplied  with  a  window function  in  STFT,  and  then  the  transform  is  computed  for  each  segment  generated. 

However, unlike STFT,  in Wavelet Transform, the width  of the  wavelet function changes with  each  spectral  component.  The  Wavelet  Transform,  at  high  frequencies,  gives  good

time  resolution  and  poor  frequency  resolution,  while  at  low  frequencies;  the  Wavelet Transform gives good frequency resolution and poor time resolution. 

2.2 The Continuous Wavelet Transform and the Wavelet Series

 The  Continuous  Wavelet  Transform  (CWT)  is  provided  by  equation  2.1,  where

x(t)  is  the  signal to be analyzed.  (t) is  the mother  wavelet  or  the  basis  function.  All  the wavelet functions used in the transformation are derived from the mother wavelet through translation (shifting) and scaling (dilation or compression). 

The mother  wavelet  used  to  generate  all  the  basis  functions  is  designed   based  on  some desired characteristics associated with that functio n. The translation par ameter   relates to

the  location  of  the  wavelet  function  as  it  is  shifted  through  the  signal.  Thus,  it corresponds  to  the  time  information  in  the  Wavelet Transform.  The  scale  parameter  s  is

defined  as |1/frequency|  and  corresponds  to  frequency  information.  Scaling  either  dilates (expands)  or  compresses  a  signal.  Large  scales  (low  frequencies)  dilate  the  signal  and

provide  detailed  information  hidden  in  the  signal,  while  small  scales  (high  frequencies) compress  the  signal  and  provide  global  information  about  the  signal.  Notice  that  the Wavelet Transform merely performs the convolution operation of the signal  and the basis

function. 

 The  above an alysis becomes  ver y  useful  as  in  most  practical  applications;  high

frequen cies  (low  scales)  do  not  last  fo r  a  long  duration,  but  instead,  appear  as  short bursts, while low frequencies (high scales) usually last for entire duration of the signal.  

The Wavelet  Series is obtained  by discretizing  CWT.  This  aids in  computation of  CWT using computers  and is  obtained by sampling  the  time-scale plane.  The  sampling rate can

be  changed  accordingly  with  scale  change  without  violating  the  Nyquist  criterion. Nyquist criterion states that,  the  minimum  sampling  rate that allows reconstruction of the original signal is  2  radians,  where   is the highest frequency in the  signal.  Therefor e, as the  scale  goes  higher  (lower  frequen cies),  the  sampling  rate  can  be  decreased  thus

reducing the number of computations. 

The Wavelet  Series is  obtained b y  discretizing CWT. This aids  in  computation of CWT  using  computers  and  is  obtained   by  sampling  the  time-scale  plane.  The  sampling rate  can  be  changed  accordingly  with  scale  change  without  violating  the  Nyquist criterion.  Nyquist  criterion  states  that,  the  minimum  sampling  rate  that  allows

reconstruction of the o riginal signal is 2  radians,  where   is the highest frequency in the signal. Therefore,  as  the  scale  goes  higher  (lower  frequencies),  the sampling  rate  can  be decreased thus reducing the number of computations. 

DISCRETE WAVELET TRANSFORM

3.1. INTRODUCTION 

 The  Wavelet  Series  is  just  a  sampled  version  of  CWT  and  its  computation  may consume significant amount  of time and  resources,  depending  on the  resolution required. The Discrete Wavelet Transform (DWT), which  is based on  sub-b and coding, is found  to yield  a  fast  computation  of  Wavelet Transform.  It  is  easy  to  implement  and  reduces  the computation time and resources required.

The foundations of  DWT go back to 1976  when  techniques to  decompose  discrete time  signals  were  devised.  Similar  work  was  done  in  speech  sign al  coding  which  was named  as  sub-band  coding.  In  1983,  a  technique  similar  to  sub-band  coding  was developed  which  was  named  pyramidal  coding.  Later  man y  improvements were made  to these coding schemes, which resulted in efficient multi-resolution analysis schemes. 

In  CWT,  the  signals  are  analyzed  using  a  set  of  basis  functions,  which  relate  to each  other  b y  simple  scaling  and  translation.  In  the  case  of  DWT,  a  time-scale representation  of  the  digital  signal  is  obtained  using  digital  filtering  techniques.  The

signal  to  be  analyzed  is  passed  through  filters  with  different  cutoff  frequencies  at different scales. 

3.2.    DWT and Filter Banks

3.2.1 Multi-Resolution Analysis using Filter Banks 

 Filters are one  of the  most  widely  used  signal processing  functions. Wavelets can be  realized  by  iteration  of  filters  with  rescaling.  The  resolution  of  the  signal,  which  is  a measure  of  the  amount  of  detail  information  in  the  signal,  is  determined  by  the  filtering operations,  and  the  scale is determined  by upsampling  and  downsampling (subsampling) operations.

The  DWT  is  computed  by  successive  lowpass  and  highpass  filtering  of  the discrete  time-domain signal as  shown in figure 3.1. This is called the Mallat algorithm  or

Mallat-tree  decomposition.  Its  significance  is  in  the  manner  it  connects  the  continuous- time  multiresolution  to  discrete-time  filters.  In  the  figure,  the  sign al  is  denoted  by  the sequence x [n], where  n  is an integer.  The low  pass filter is denoted by  G0 while  the  high pass  filter  is  denoted  by  H0.  At  each  level,  the  high  pass  filter  produces  detail information;  d[n],  while  the  low  pass  filter  associated  with  scaling  function  produces coarse  approximations, a[n]. 

                       At each decomposition level, the half  band filters produce signals  spanning only  half  the frequency  band.  This doubles the frequency  resolution as  the uncertainty  in

frequen cy is reduced by half. In accord ance  with Nyquist’s rule if the original signal has a  highest frequency  of  , which  requires a sampling frequency of  2   radians,  then  it now

has a highest frequency of  /2 radians.  It can now  be sampled at a frequency of   radians thus discarding half the samples with  no loss of information. This decimation by 2 halves

the  time  resolution  as  the  entire  signal  is  now  represented  by  only  half  the  number  of samples. Thus, while the half band  low pass filtering removes half  of the frequencies  and

thus halves the resolution, the decimation by 2 doubles the scale. 

With  this  approach,  th e  time  resolution  becomes  arbitrarily  good  at  high frequencies,  while the  frequency  resolution becomes  arbitrarily  good at  low  frequencies.The  filtering  and  decimation  process  is  continued  until  the desired  level  is  reached.  The  maximum number of levels depends  on the length of the signal. The DWT  of the  original signal is  then  obtained  by  concatenating all  the  coefficients,  a[n]  and  d[n],  starting  from the last level of decomposition.

Figure 3.2 shows  the reconstruction of the original signal from the wavelet

coefficients.  Basically,  the  reconstruction  is  the  reverse  process  of  decomposition.  The approximation  and  detail  coefficients  at  every  level  are  upsampled  by  two,  passed through  the  low  pass  and  high  pass  synthesis  filters  and  then  added.  This  process  is

continued  through  the  same  number  of  levels  as  in  the  decomposition  pr ocess  to  obtain the original signal. The Mallat algorithm works equally well if  the  analysis  filters,  G0 and H0, are exchanged with the synthesis filters,  G1  & H1. 

3.2.2 Conditions for Perfect Reconst ruction 

 In  most  Wavelet  Transform applications,  it  is  required that  the  origin al  signal b e synthesized  from  the  wavelet  coefficients. To achieve  perfect  reconstruction  the analysis and  synthesis  filters  have  to  satisfy  certain  conditions.  Let  G0(z)  and  G1(z)  be  the  low pass analysis  and synthesis filters, respectively and H0(z) and H1(z) the high  pass analysis and  synthesis  filters  respectively.  Then  the  filters  have  to  satisfy  the  following  two

conditions:

 The first  condition  implies  that  the  reconstruction  is  aliasin g-free  and  th e  second

condition  implies  that  the  amplitude  distortion  has  amplitude  of  one.  It  can  be  observed that  the  perfect  reconstruction  condition  does  not  change  if  we  switch  the  analysis  and synthesis filters. 

There  are  a  number  of  filters,  which  satisfy  these conditions.  But not  all of  them

give  accurate  Wavelet  Transforms,  especially  when  the  filter  coefficients  are  quantized.The  accuracy  of  the  Wavelet  Transform  can  be  determined  after  reconstruction  by

calculating the Signal  to  Noise Ratio  (SNR)  of the signal.  Some applications like  pattern recognition  do  not  need  reconstruction,  and  in  such  applications,  the  above  conditions

need not apply. 

3.2.3 Classification of wavelets 

 We  can  classify  wavelets  into  two  classes: 

 (a)  orthogonal  and  (b)  biorthogonal.

Based on the application, either of them can be used. 

(a) Features of orthogonal wavelet filter banks 

 The coefficients of  orthogonal filters are r eal  numbers.  The  filters are of the same length  and  are  not  symmetric.  The  low  pass  filter,  G0  and  the  high  pass  filter,  H0  are related to each other b y  

                                                             

                                               H0 (z) = z -N G0 (-z-1)…………………………………...(3.3 )

The two  filters are alternated  flip  of each  other.  The  alternating  flip  automatically gives  double-shift orthogonality  between the  lowpass  and  highpass  filters, i.e.,  the  scalar product  of  the  filters,  for  a  shift  by  two  is  zero.  i.e.,  G[k]  H[k-2l]  =  0,  where  k,l Z. Filters  that  satisf y  equation  3.3  are  known  as  Conjugate  Mirror  Filters  (CMF).  Perfect reconstruction is possible with alternating flip. 

Also,  for  perfect  r econstruction,  the  synthesis  filters  are  identical  to  the  analysis filters  except  for  a  time  reversal.  Orthogonal  filters  offer  a  high  number  of  vanishing moments. This property is useful in many signal and image processing applications. They have regular structure, which leads to easy implementation and scalable  architecture.

  

(b)Features of biorthogonal wavelet filter banks 

In  the  case  of  the  biorthogonal  wavelet  filters,  the  low  pass  and  the  high  pass filters  do  not  have  the  same  length.  The  low  pass  filter  is  always  symmetric,  while  the high  pass  filter  could  be  either  symmetric  or  anti-symmetric.  The  coefficients  of  the filters are  either real numbers or integers. 

For perfect r econstruction,  biorthogonal  filter  bank has  all  odd  length  or all even length  filters.  The  two  analysis  filters  can  be  symmetric  with  odd  length  or   one symmetric  and  the  other  antisymmetric  with  even  length.  Also,  the  two  sets  of  analysis and  synthesis  filters  mu st  be  dual.  The  linear  phase  biorthogonal  filters  are  the  most popular filters for d ata compression applications. 

3.3 Wavelet Families 

            There  are  a  number  of  basis  functions  that  can  be  used  as  the  mother wavelet  for  Wavelet  Transformation.  Since  the  mother  wavelet  produces  all  wavelet functions  used  in  the  transformation  through  translation  and  scaling,  it  determines  the characteristics of the resulting Wavelet  Transform.  Therefore, the details  of the particular

application  should  be  taken  into  account  and  the  appropriate  mother  wavelet  should  be chosen in order to use the Wavelet Transform effectively. 

Figure 3.3 illustrates some of the  commonly used wavelet functions. Haar  wavelet is  one  of  the  oldest  and  simplest  wavelet.  Therefore,  any  discussion  of  wavelets  starts with  the  Haar  wavelet.  Daubechies  wavelets  are  the  most  popular  wavelets.  They represent  the  foundations  of  wavelet  signal  processin g  and  ar e  used  in  numerous applications.  These  are  also  called  Maxflat  wavelets  as  their  fr equency  responses  have maximum  flatness  at  frequencies  0  and  .  This  is  a  very  desirable  property  in  some applications.  The  Haar,   Daubechies,  Symlets  and  Coiflets  are  compactly  supported orthogonal  wavelets.  These  wavelets  along  with  Meyer  wavelets  are  capable  of  perfect reconstruction.  The  Meyer,  Morlet  and  Mexican  Hat  wavelets  are  symmetric  in  shape. The wavelets  ar e  chosen  based  on  their  shape  and  their  ability  to  analyze  the  signal  in  a particular application. 

METHODS OF EYE DETECTION

4.1 INTRODUCTION:

                               A lot of research work has been published in the field of eye detection in the last decade. Variou s techniques have been proposed using texture,  depth, shape and

color  information  or  combinations  of  these  for  eye  detection.  Vezhnevets  focus  on several  landmark  points  (eye  corners,  iris  border  points),  from  which  the  approximate

eyelid  contours are estimated. The upper  eyelid points  are found  using on the observation that  eye  border  pixels  are  significantly  darker  than  surrounding  skin  and  sclera.  The

detected  eye  bound ary  points  are  filter ed  to  remove  outliers  and  a  polynomial  curve  is fitted  to  the  remaining  boundary  points.  The  lower  lid  is  estimated   from  the  known  iris and eye. Some of the famous eye detection techniques are discussed below.

4.2 TEMPLATE MATCHING METHOD: 

                               Reinders  present  a  method  where  based  on  the  techniq ue  of  template matching  the  positions  of  the  eyes  on  the  face  image  can  be  followed  throughout  a sequence  of  video  images.  Template  matching  is  one  of  the  most  typical  techniques  for

feature extraction. C orrelation is commonly  exploited to measure the  similarity         between  a stored  template  and  the  window  image  under  consideration.  Templates  should  be

deliberately designed  to  cover variety  of possible  image  variations.  During  the  search  in the  whole  image,  scale  and  rotation  should  also  be  considered  carefully  to  speed  up  the process.  To  increase  the  robustness  of  the  tracking  scheme  the  method  automatically

generates a codebook of images representing the  encountered different  appearances of the eyes.  Yuille  first  proposed  using  deformable  templates  in  locating  human  eye.  Th e

weaknesses  of  the  deformable  templates  are  that  the  processing  time  is  lengthy  and success  relies  on  the  initial  position  of  the  template.  Lam  introduced  the  concept  of  eye

corners to improve the deformable template appro ach. 

4.3 USING PROJECTION FUNCTION:

Saber  and  Jeng  proposed  to  use  facial  features  geometrical

structure  to  estimate  the  location  of  eyes.  Takacs  developed  iconic  filter  banks  for detecting  facial  landmarks.  projection  functions  have  also  been  employed  to  locate  eye

windows. Feng and Yeun developed a variance projection function for locating the corner points  of the  eye.  Zhou and Geng propose  a h ybrid projection function to  locate  the  eyes.

By  combining  an  integral  projection  fun ction,  which  considers  mean  of  intensity,  and  a variance  projection  function,  which  considers  the  variance  of  intensity,  the  hybrid function  better  captures  the  vertical  variation  in  intensity  of  the  eyes.  Kumar  suggest  a

technique  in  which  possible  eye  areas  are  localized  using  a  simple  thresholding  in  color space  followed  b y  a  connected  component  analysis  to  quantif y  spatially  connected

regions  and  further  reduce  the  search  space  to  determine  the  contending  eye  pair windows.  Finally the  mean and  variance projection  functions are utilized  in each eye pair

window  to  validate  the  presence  of  the  eye.  Feng  and  Yeun  emplo y  multi  cues  for  eye detection on gray images using variance projection function.

4.4 IR METHOD:

                  The most common  approach employed to achieve  eye  detection in  real-time is by  using  infrared  lighting  to  capture  the  physiological  properties  of  eyes  and  an

appearance-based  model  to  represent  the  eye  patterns.  The  appearance-based  approach detects eyes based on the intensity distribution of the eyes by exploiting the differences in

appearance of eyes from the rest of the face. This method requires a significant number of training  data  to  enumerate  all  possible  appearances  of  eyes  i.e.  representing  the  eyes  of different  subjects,  under different face orientations,  and diff erent  illumination conditions.

The  collected  data  is  used  to  train  a  classifier  such  as  a  neural  net  or  support  vector machine to achieve detection.

4.5 SUPPORT VECTOR MACHINES (SVMs). 

          Support  Vector  Machines  (SVMs)  have  been  recently  proposed

by Vapnik and his co-workers as a very effective method for general-purpose pattern recognition. Intuitively,  given  a  set  of  points  belonging  to  two  classes,  a  SVM  finds  the  hyper-plane  that separates  the  largest  possible  fraction  of  points  of  the  same  class  to  the  same  side  while maximizing the distances from either class to the hyper-plane. This hyper-plane is called Optimal

Separating  Hyper-plane  (OSH).  It  minimizes  the  risk  of  misclassifying  not  only  the  samples  in the training set  but also the  unseen samples in the  test set. The  application of SVMs  to computer vision area has emerged recently. Osuna train a SVM for face detection, where the discrimination is  between two classes: face and  non-face,  each  with  thousands of  samples.  Guo and  Stan  show that the  SVMs  can  be  effectively  trained  for  face  recognition  and  is  a  better learning  algorithm than the nearest center approach.

Graph  Matching.  After  all  images,  including  the  gallery  images  and  the  probe  images,  are

extracted using EBGM procedure, the faces are represented as labelled face graphs. The matching procedure then involves the distance computation of the jets between different graphs, which is

4.6 Hidden Markov Models (HMMs):

HMMs  are  generally  used  for  the  statistical  modelling  of  non-

stationary vector time series. By considering the facial configurable information as a time varying sequence,  HMMs  can  be  applied  to  face  recognition.  The  most  significant  facial  features  of  a frontal  face image,  including  the  hair,  forehead, eyes, nose  and mouth,  occur in  a natural  order from  top to bottom,  even  if the image has  small  rotations  in  the  image  plane, and/or  rotations  in the plane perpendicular to the image plane. Based on this observation, the image of a face may be modeled using  a one-dimensional HMM by  assigning  each  of these regions a  state as 

                      

 Given  a  face  image  for  one subject  in  the  trainin g set, the  goal  of the  training  stage  is  to

optimize  the  parameters  to  best  describe  the  observation.  Recognition  is  carried  out  by matching the  test image against  each  of  the  trained models.  To  complete  this  procedure,

the  image  is  converted  to  an  observation  sequence  and  the  likelihood  is  computed  for each  stored  model.  The  model  with  the  highest  likelihood  reveals  the  identity  of  the unknown  face.  The  HMM  approach  has  shown  the  ability  to  yield  satisfactory

recognition  rates.  However,  HMMs  are  processor  intensive  models,  which  implies  that the algorithm may run slowly. The HMM lead to the efficient detection of  eye strips.

4.7 WAVELET BASED METHOD:

                        

 Our  project  is  based  on  this  method  of  eye  detection.  Wavelet

decomposition  provides  local  information  in  both  space  domain  and  frequency  domain. Despite  the  equal  subband  sizes,  different  subbands  carry  different  amounts  of

information.  The  letter  ‘L’  stands  fo r  low  frequency  and  the  letter  ‘H’  stands  for  high frequen cy.  The  left  upper  band  is  called  LL  band  because  it  contains  low  frequency

information  in  both  the  row  and  column  directions.  The  LL  band  is  a  coarser

approximation  to  the  original  image  containing  the  overall  information  about  the  whole image.  The  LH  subband is  the  result  of applying the filter  bank column  wise and  extracts

the  facial  features  v ery  well.  The  HL  subband,  which  is  the  result  of  applying  the  filter bank  row  wise,  extracts  the  outline  of  the  face  boundar y  very  well.  While  the  HH  band shows  the  high  frequency  component  of  the  image  in  non-horizontal,  non-vertical

directions  it  proved to be a  redundant subband and was not  considered having significant

information  about  the  face.  This  observation  was  made  at  all  resolutions  of  the  image.This  is  the  first  level  decomposition.  Finally a  fixed no.  of   maximum peaks are selected from  LH subband,  which are fed as inputs to the neural n etwork back propagation model

or  RBF  or  neuro-fuzzy  model  is  used  to  train  that  required  network.  According  to  the outputs  of  those  peaks,  after  being  passed  through  the  updated  weight  and  bias  values, they are  categorized into eye  parts and  non-eye parts.  our project  is based on this method

of eye detection.

WAVELET BASED METHOD FOR EYE DETECTION 

5.1 INTRODUCTION:

                  

                     The system consists mainly of two stages training and detection stage. A

block diagram of these two stages is shown in Figure 1.

 5.2 Acquisition of Training Data:

                                    The  training  data  typically  consists  of  50  images  of  different persons  with  different  hairstyles,  different  illumination  conditions  and  varying  facial

expressions. Some of the images have different states of the eye such as  eyes closed. The size of the images varies from 64x64 to 256x256.

5.3 Discrete Wavelet Transform:

                                 Wavelet  decomposition  provides  local  information  in  both  space domain and frequency domain.  Despite  the equal subband  sizes, different subbands carry

different  amounts  of  information.  The  letter  ‘L’  stands  for  low  frequency  and  the  letter ‘H’stands  for high frequency.  The left  upper  band  is  called  LL  band  because  it  contains low  frequency  info rmation  in  both  the  row  and  column  directions.  The  LL  band  is  a

coarser app roximation  to  the original  image  containing the  overall  information about the whole  image.  The  LH  subband  is the  result of  applying  the  filter  bank  column  wise  and extracts the facial features very  well. The HL subband,  which is the result of applying the 

filter  bank  row  wise,  extracts the  outline  of  the  face  boundary  very  well.  While  the  HH band  shows  the  high  frequency  component  of  th e  image  in  non-horizontal,  non-vertical directions  it  proved to be a  redundant subband and was not  considered having significant

information  about  the  face.  This  observation  was  made  at  all  resolutions  of  the  image.This  is  the  first  level  decomposition.  A  CDF  (2,  2)  biorthogonal  wavelet  is  used.  Gabor Wavelets  seem  to be  the  most  probable  candidate  for  feature  ex traction.  But  they  suffer

from  certain  limitations  i.e.  they  cannot  be  implemented  using  Lifting  Scheme  and secondly  th e Gabor  Wavelets form a non-orthogonal  set  thus making  the computation  of

wavelet  co efficients  difficult  and  expensive.  Special  hardware  is  required  to  make  the algorithm work  in  real time. Thus  choosing  a  wavelet  for  eye  detection  depends  on a lot of  trial  and  error.  Discrete  Wavelet Transform  is  recursively  applied  to  all the  images  in

the training data set until the lowest frequency subband is  of size 32x 32 pixels  i.e. the LH subband  at  a  particular  level  or  depth  of  DWT  is  of  size  32x32.  The  original  image’s

grayscale  image is shown  in  figure  5.2. Th e LH  subband at resolution  32x32 is  shown  in Figure 5.3.Here we have used HAAR  wavelet instead of  Gabor wavelet while calculating

wavelet transform.

fig-5.3 the LH sub-band figure 

We  take  the  modulus  of  the  wavelet  coefficients  in  the  LH  subband.  Experiments  were performed  to  go  to  a  resolution  even coarser  than 32x32.  However,  it  was  observed  that in  certain  cases  the  features  would  be  too  close  to  each  other  and  it  was  difficult  even manually too to  separate them. This would burden the Neural Network model and  a small

error in locating the eyes at this low resolution would result in a large error in locating the eyes in the original image.

5.4 Detection of Wavelet Maxima:

                                          Our approach to eye  detection is based on the observation that, in  intensity  images  eyes  differ  from  the  rest  of  the  face  because  of  their  low  intensity. Even if the eyes are closed, the  darkness of the eye sockets is sufficient  to extract the eye

regions.  These  intensity  peaks  are  well  captured  by  the  wavelet  coefficients.  Thus, wavelet  coefficients  have  a  high  value  at the  coordinates  surrounding  the  eyes.  We  then

detect the wavelet maxima  or the wavelet peaks  in  this  LH subband  of  resolution 32x32. Note  that  several  such  peaks  are  detected,  which  can  be  the  potential  lo cations  of  th e

eyes. The intensity peaks are shown in Figure 5.4 and 5.5.

Fig-5.5 LH sub-band with peaks replaced by its 3 *3 neighborhood wavelet coefficient 

5.5 Neural Network Training:

    The wavelet  peaks detected  are  the  center of  potential eye windows.  We then feed 3x3 neighborhood  wavelet  coefficients  of  each  of  these  local  maxima’s  in  32x32  LH

subbands  of  all  training  images  to  a  Neural  Network  for  training.  The  Neural  Network has 9 input nodes,  4 hidden nodes, and 2 output nodes. A diagram of the Neural Network

architecture is shown in Figu re 5.6.. A (1,-1) at  the output of Neural Network indicates  an eye  at  the  location  of  the  wavelet  maxima  whereas  (-1,  1)  indicates  a  non-eye.  Two output  nodes  instead  of  one  were  taken  to  improve  the  performance  of  the  Neural

Network.  MATLAB’s  Neural  Network  Toolbox  was  used  for  simulation  of  the  back propagation  Neural  Network. A conjugate  gr adient learning rate of  0.4 was  chosen while

training. This completes the training stages for neural networks b ack propagation model.

Here we hav e used the MLP (multi-layer p erceptions) back-propagation model for neural network training. It consists of having 9 neurons in the input layer. In the hidden layer or 2nd layer has 5 neuron for processing .In the 3rd or output layer h as the two nodes for

showing the output. We have taken two output node insists of one to get  a better accuracy towards detecting eye. After this you have to find  eye part & non-eye part in the figure

from neural network model. Where an output of  (1, -1) indicates th e presence of an eye & output of (-1,1) indicates the presence of a non-eye.

 EXPERIMENTATION & RESULTS 

7.1 SOURCE CODE 

Prog ram no.1   

 

(Wavelet transform part)

 function[inp]=aisT(a)

% clear all;

% close all;

% clc;

dwtmode('zpd');

% a=double(rgb2gray(imread('C:\MATLAB7\work\project\ais2.jpg') ));

 imshow(uint8(a))

 %figure

sizea=size(a)

[c,s]=  wavedec2(a,4,'haar');

               % ap_cf  = appcoef2(c,s,'h aar',4);

               % sizeAP_CF=size(ap_cf)

sizec=size(c)

h_cf2 = detcoef2('h ',c,s,4); 

[m,n]=size(h_cf2);       %n used later

sizehcf=size(h_cf2)

% sizevcf=size(v_cf2)

% sizedcf=size(d_cf2)

% imshow(uint8(ap_cf))               %for LL image

 %figure                                   %figure of LH

%imshow(uint8(h_cf2))

imview(h_cf2)

% % figure

% % imshow(uint8(v_cf2))

% % figure

% % imshow(uint8(d_cf2))

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%%%%%%          DETECTION  OF WAVELET

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%%%%%%                        MAXIMA

% LH=abs(h_cf2);

LH1=zeros(m*n,1);            % LH1 just used for descending matrix h_cf....

%^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

k=1;

for i=1:m

    for j=1:n

        LH1(k)= abs(h_cf2(i,j));       %(h_cf2(i,j));     

        k=k+1;

    end

end

   LH=sort(LH1,'descend ');                     %all values h_cf2 r stored in LH1

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%max 

peak wav elet

ln=25;

max_pk=zeros(ln,1);

for i=1:ln

    max_pk(i)=LH(i);                                        %the value of top  ln=25 max wave coef.. in LH

figure(h_cf2)

end

% % % % % % % % % % % % % % % % % % % % % % % % % % % % % % 

dhcf2=zeros(m,n);

for i=1:m

    for j=1:n

        if any(  abs(h_ cf2(i,j))>=max_pk )     %  (h_cf2(i,j))>=max_pk )

            dhcf2(i,j)=h_cf2(i,j);                                             % dchf  is used  for  max  peak  detectiopn                              

&                                                                                                      others zero

        end

    end

end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%%%%%%

% dhcf2

imview(dhcf2)

%    figure

%    imshow(uint8(dhcf2))

 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%  for            

neighbourhood configuration

   inp=zeros(ln,13);

   size(inp)

   l=1;k=1;

   ph_cf2=[zeros(1,n+2);zeros(m,1),h_cf2,zeros(m,1);zeros(1,n+2)];      %2  avoid  error

Attempted to access h_ cf2(13,51); 

   pdhcf2=[zeros(1,n+2);zeros(m,1),dhcf2,zeros(m,1);zeros(1,n+2)];          %index  out  of

bounds because size(h_cf2)=[38,50].

   

   [size_LHx,size_LHy]=size(h_cf2);

   

   for i=2:m+1

      for j=2:n+1

        if dhcf2(i-1,j-1)~=0 

            

            pdhcf2(i-1,j-1)=ph_cf2(i-1,j-1);                  p dhcf2(i-1,j) =ph_cf2(i-1,j);        pdhcf2(i-

1,j+1)=ph_cf2(i-1,j+1);

            pdhcf2(i,j-1)=ph_cf2(i,j-1);                          pdhcf2(i,j)=ph_cf2(i,j);       

pdhcf2(i,j+1)=ph_cf2(i,j+1);

            pdhcf2(i+1,j-1)=p h_cf2(i+1,j-1);                  pdhcf2(i+1,j)=ph_cf2(i+1,j);  

pdhcf2(i+1,j+1)=ph_cf2(i+1,j+1);

            

              inp(l,k)   =ph_cf2(i-1,j-1);              inp(l,k+1)=ph_cf2(i-1,j);      inp(l,k+2)=ph_cf2(i-

1,j+1); 

              inp(l,k+3)=ph_cf2(i,j-1);                      inp(l,k+4)=ph_cf2(i,j);     

inp(l,k+5)=ph_cf2(i,j+1);

              inp(l,k+6)=ph_cf2(i+1,j-1);                  inp(l,k+7)=ph_cf2(i+1,j);   

inp(l,k+8)=ph_cf2(i+1,j+1);

%               inp(l,k+10)=i-1/size_LHx ;       inp(l,k+9)=j-1/size_LHy;

                   inp(l,k+10)=i-1;       inp(l,k+9)=j-1;

             l=l+1;     k=1;

                   

        end

      end

   end

   inp(:,10:11)

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%%  

   for i=1:m

    for j=1:n

        xdhcf2(i,j)=pdhcf2(i+1,j+1);

    end

   end

   imview(xdhcf2)

%     figure

 %       imshow(uint8(xdhcf2))

% max_pk

% pdhcf2

% ph_cf2

Prog ram no.2

 

(Neural network model part)

%for aiswariya 

close all;

clear all;

clc;

a=double(rgb2gray(imread('C :\MATLAB7\work\project\ais2.jpg')));

pa=aisX(a);

pk=pa(:,1:9);

p=pk';

 Tp=[1 1 1 1 1 1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1];

 T=[Tp;-Tp];

 net = newff( minmax(p),[5 2],{'tansig' 'purelin '});

 Y = sim(net,p);

%  plot(P,T,P,Y,'o');

%   net.trainParam.show  = 50;

 net.trainParam.lr = 0.4;

 net.trainParam.epo chs = 4000;

 net.trainParam.goal = 1e-5;

 [net,tr] = train(net,p,T);

 Y = sim(net,p)

%   plot(P,T,P,Y,'o')

%   Y=sim(net,3.5)

 b=double(rgb2gray(imread('C:\MATLAB7 \work \project\ais6.jpg')));

 pb=aisT(b);

 pbc=pb(:,1:9);

 Y=sim(net,pbc')

Y1(output for training image) =

  Columns 1 through 7 

    0.9090    0.9173    0.9187    0.8910    0.9654    0.9287   -0.8449

   -0.9604   -0.9515    -0.8977   -1.1043   -0.9719   -0.9439    0.7588

  Columns 8 through 14 

   -0.9669   -0.8692   -0.9752   -0.9505   -0.9779   -0.9481    -0.0072

    0.8855    0.9132    0.7044    0.8979    0.7072    0.9427       0.8075

  Columns 15 through 21 

    -0.9867   -0.8649   -0.9705   -0.9671   -0.9180   -0.8279    -0.9568

    0.9640    0.9966    0.9099    0.8198    0.8876    0.9275    0.9610

  Columns 22 through 25 

   -0.9707    -0.9030   -0.9872   -0.9412

    0.9744    0.9312    0.9944    0.9187

Y2(output for test image)=

  Columns 1 through 7 

     -0.9593    -0.8823      0.9363    0.9128      0.8900     0.9825      0.9067

      0.8770       0.9303    -0.9011   -0.9900    -0.9881    -0.9614    - 0.9667

  Columns 8 through 14  

     0.8744    -0.8321    -0.8557       -0.9644    -0.9089   -0.8669   -0.9586

   -0.9106      0.9858     0.9368        0.9782      0.8298    0.9175     0.8018

  Columns 15 through 21 

   -0.9604    -0.8081   -0.8315   -0.8783    -0.8116    -0.9650    -0.9011

    0.8948      0.9015    0.9271    0.9845      0.9891     0.9533      0.8772

  Columns 22 through 25 

   -0.8403   -0.8179   -0.9558    -0.9443

    0.9556    0.9958     0.8170      1.0770

CONCLUSION

7.1. Performance

A  number  of  experiments  were  done  to  test  the  robustness  of  the  algorithm  and  to increase  the  accuracy  of  eye  detection.  Various  architectures  of  Neural  Networks  with

different  learning  rates  were  tried  and it  was  found  that back propagation with conjugate gradient learning seemed  to be the best choice.  A very high learning rate of 0.8 was

Chosen because the learning algorithm was getting trapped in local minima while training the  network.  Final  training  was  stopped  when  the  error  graph,  as  shown  in  Figure  7.1,

didn’t show any significant fluctuation. 

An  ex periment  was  done  in  which  the  face  was  analyzed  using  wavelet  packets  and  it was found that  most of  the  information  was retained by the low  frequency sub bands and the  high  frequency  packets  had  no  information.  Images  with  different  states  of  the  eye

(closed, open,  half  open,  looking  sideways,  head  tilted  etc.) and  varying  eye  width  were chosen. The eye positions found were compared  with the positions that were pointed out

manually. The eyes  were  correctly  located when  its  location  is within two  pixels, in  both x  and  y  directions,  of  the  manually  assigned  point.  The  variation  of  2  pix els  is

deliberately  allowed,  to  compensate  for  the  inaccuracies  in  the  location  of  eyes  during training.  An  accuracy  of  88%  was  observed  in  the  final  location  of  the  eyes. A  database

of  60 test  images  was  evaluated  for performan ce.  All  these test  images  were  captured  in totally  different environment  conditions  and  wer e  not  included  while  training  the Neural Network.  Most  of  the  error  cases  occurred  in  images  with  complex  background.  Also

there was an  error in accurately  determining the exact  location  of the eyes since a 1-pixel shift  at  a  resolution  of  32x32  corresponded  to  a  larger  shift  in  the  exact  location  of  the

eye.  In  some  cases  the  Neural  Network  classified only  1  peak  as  an  eye  in  spite  of  the

presence of  2 eyes  in the image. In a few  cases observations were  made in which  regions of  the  face  not  belonging  to  the  eyes  were  d etected  as  eyes.  In  other  cases  more  than  2 eyes  were  indicated  in  the  image.  In  contrast,  the  perfo rmance  of  this  algorithm,  which

uses  wavelets  as  a  preprocessor  to  Neural  Networks,  the  algorithm  with  only  Neural Networks, achieved an accuracy of 81% in d etecting the exact location of the eyes.

7.2 CONCLUSION:

This  type  of  approach  gives  a  n ew  dimension  to  the  existing  eye  detection  algorithms. The present  algorithm  is  robust and at  par  with  the  other existing  methods but still  has  a lot of scope for improvement. In this type of approach a wavelet subband approach in 

using Neural Networks  for eye detection. Wavelet Transform is adopted to decompose  an image  into  different  sub bands  with  different  frequency  compon ents.  A  low  frequency subband  is  selected  for  feature  extraction.  The  proposed  method  is  robust  against

illumination, background, facial expression ch anges and also works for images of

different sizes. However,  a combination of information in different frequency bands at different  scales,  or usin g multiple  cues can even  give  better  performance.  Further  studies

                                   in using Fuzzy Logic fo r data fusion of multiple cues will give better results.

REFERENCES:

1. M.J.T. Reinders, "Eye tracking by template matching using an  automatic codebook

generation scheme", Third  annual  conferen ce of the Advanced School for C omputing and Imaging, pp. 85-91, Heijen, The Netherlands, June 1997.

2.  Kumar,  Thilak  R  and  Raja,  Kumar  S  and  Ramakrishnan,  “Eye  d etection  using  color cues  and  projection  functions”,  Proceedings  2002  International  Conference  on  Image

Processing ICIP, pages Vol.3 337-340, Rochester, New York, USA.

3. K. M. Lam, H. Yan, ” Locating and extracting the eye in human f ace images”, Pattern Recognition, Vol. 29, No. 5   pp.771-779.(1996)

4.  Takacs,  B.,  Wechsler,  H.,  "Detection  of f aces  and  facial  landmarks  using iconic  filter banks", Pattern Reco gnition, Vol. 30, No. 10, Octo ber 1997, pp. 1623-1636.

5. Vezhnevets V., Degtiareva A., "Robust and Accurate Eye Contour Extraction", Proc. Graphicon-2003, pp. 81 -84, Moscow, Russia, September 2003.

6.  Erik  Hjelms  and  Jrn  Wroldsen,  "Recognizing  faces  from  the  eyes  only",  Proceedings of the 11th Scandinavian Conference on Image An alysis, 1999

7.  A.  Pentland,  B.  Moghaddam,  T.  Starner,“View-based  and  modular  eigenspaces  for face recognition”, Proceedings of the IEEE Intern ational Conference  on Computer Vision and Pattern Recognition, Seattle, WA, 1994, pp.84-91.

8. C. Morimoto, D. Koons, A. Amir, and M.Flickner, “Real-Time Detection of Eyes and Faces”,  Proceedings  of  1998  Workshop  on  Perceptual  User  Interfaces,  pages  117-120,

San Francisco, CA, November 1998.

9.  W.  Sweldens  and  P.  Schrder,  "Building  your  own  wavelets  at  home",  Wavelets  in Computer Graphics, pp. 15--87, ACM SIGGRAPH Course notes, 1996.

10.Baback  Moghadd am  and  Ming-Hsuan  Yang.  Gender  Classification  with  Support Vector  Machine,   Proceeding  of  the  4th  International  Conference  on  Face  and  Gesture

Recognition, pp306-311, Grenoble, France, 2000.

11.  L.  Ma, Y.  Wan g,  T. Tan.  Iris r ecognition  using  circular  symmetric filters.      National Labo ratory  of  Pattern  Recognition,  Institute  of  Automation,  Chinese  Academy  of

Sciences, 2002.  

12. J. M. Shapiro, "Embedded image coding using zerotrees of wavelet coefficients",

IEEE Trans. on Signal Processing, Vol. 41, No. 12, pp. 3445-3463, Dec. 1993