Pattern Recognition Projects

1. MRF Models of Faces

The spatial distribution of gray level  intensities in an image can be naturally modeled using Markov Random   Field   (MRF)   models.  We   develop   and   investigate   the   performance   of   face  detection algorithms derived from MRF considerations.  For enhanced detection,  the MRF models are defined for every permutation of site indices in the image. We find the optimal  permutation that provides maximum discriminatory power to identify faces from nonfaces. The MRF models successfully detect

faces in a number of test images in real time.

2. Clustering and Feature Selection

This work proposes an unsupervised algorithm for learning  a finite mixture model from multivariate data. The adjective "unsupervised"  is justified by two properties of the algorithm:

 (i) it is capable of selecting  the number of components,  and  (ii)  unlike the  standard  expectation-maximization  (EM) algorithm,it   does   not   require   careful   initialization. 

  The   proposed   method   also   avoids   another drawback  of EM for mixture fitting: the possibility of convergence towards a singular estimate at the boundary  of  the  parameter  space.  The  novelty  of  our  approach  is that we do not use  a model selection   criterion   to   choose   one   among   a   set   of  preestimated   candidate   models;   instead,   we seamlessly  integrate  estimation and model  selection  in a  single  algorithm.  Our  technique  can be applied to any type of parametric mixture model for which it is possible to write an EM algorithm. In our  first  paper,  we illustrate it  with  experiments  involving  Gaussian  mixtures.  These  experiments testify for the  good performance of our approach.  This  approach is extended  to perform  feature selection -- the selection of "good" variables -- for learning a mixture model in an unsupervised setting. Feature selection in unsupervised learning  is much more difficult than its counter-part in supervised learning  because of the lack  of class labels.  By treating the relevance of each feature as a Bernoulli random variable, we obtain an EM algorithm that estimate both the number of components and the importance of the features simultaneously.  A complimentary approach based on a "wrapper" on the standard   EM   mixture   learning   algorithm   is  also   proposed   for  feature  selection   in   unsupervised learning.  The optimal  feature subset size is determined automatically by the entropy of assignment, instead of manually adjusted.

3. Face Modeling for Recognition

3D  Human  face   models  have  been  widely  used   in  applications  such  as   face  recognition,  facial expression   recognition,   human   action   recognition,   head   tracking,   facial   animation,   video compression/coding, and augmented reality.  Modeling  human faces provides a potential solution to the variations encountered on human face images. We propose a method of modeling human faces based   on  a   generic  face  model  (a  triangular  mesh  model)   and  individual  facial  measurements containing both shape and texture information. The modeling method adapts a generic face model to the given facial features, extracted from registered range and color images, in a global-to-local fashion. It iteratively moves the vertices of the mesh model  to smoothen the non-feature areas,  and uses the 2.5D active  contours to refine feature boundaries. The resultant face  model has been shown to be visually similar to the true  face.  Initial  results  show  that the  constructed  model  is quite useful for recognizing profile views.

4. Face Detection in Color Images

 Human  face   detection   is   often  the   first  step   in  applications  such  as  video  surveillance,  human computer interface, face recognition, and image database management. We propose a face detection algorithm   for   color   images   in   the   presence   of   varying   lighting   conditions  as  well   as   complex backgrounds.   Our   method   detects  skin   regions   over   the   entire   image,   and   then   generates   face candidates  based on the spatial  arrangement of  these skin patches.  The algorithm constructs eye, mouth,  and  boundary maps  for  verifying  each  face  candidate.  Experimental  results  demonstrate successful detection over a wide variety of facial variations in color, position, scale, rotation, pose, and expression from several photo collections.

5. Video Mosaicking (DIP)

6. Fingerprint Mosaicking

It has been observed that the reduced contact area offered by solid-state fingerprint sensors do not provide sufficient information (e.g.,  minutiae)  for high accuracy user verification. Further,  multiple impressions of the same  finger acquired by these sensors,  may have only a small  region of overlap thereby affecting the matching  performance of the verification system. To deal with this problem,  we suggest a fingerprint mosaicking scheme that constructs a composite fingerprint image using  multiple impressions.  In the proposed algorithm,  two impressions of  a finger are  initially aligned using  the corresponding  minutiae  points.  This   alignment  is  used  by  the  well-known  iterative  closest  point algorithm (ICP) to compute a transformation matrix that defines the spatial relationship between the 

two impressions.

 The transformation matrix is used in two ways:

 (a) the two impressions are stitched together to generate a composite image.  Minutiae points are then detected in this composite image. 

(b) the minutia maps obtained from  each of  the individual  impressions are  integrated to create a larger  minutia  map.  The  availability  of   a  composite   template  improves  the   performance  of  the  fingerprint matching system as is demonstrated in our experiments.