Machine Recognition in Complex Domain - High Dimensional Neurocomputing: Growth, Appraisal and Applications - page 131

7.1.1.1 Structural or Model-Based Approach
It is based on extracting structural features that are local structure of images. It uses
geometric relationships among these extracted structures. One of the popular meth-
ods of facial feature extraction is finding the geometry of face characteristics elements
such as face contour, mouth, nose, eyes, etc. An explicit modeling of face variations
has intrinsic physical relationship with real faces. Such model-based feature extrac-
tion (deformable templates, active contours, etc) is complicated and laborious. It is
model fitting process and recognition results highly depend on the fitting results. The
location and local statistics are fed into a structural classifier. The geometric features
such as eyebrow thickness, nose vertical position and width, chin shape, etc, are
extensively used for matching. Though, this approach is less sensitive to variations
in illumination and to irrelevant information on an image; however, this approach
of feature extraction is not much reliable. It has difficulties when appearances of
features change significantly. For example, a smiling face and a frowning face are
considered as two totally different image templates in rigid body sense. Similarly,
faces with open eyes, close eyes and eyes with glasses. These techniques utilize pure
geometry methods and require a lot of mathematics [ 14 ], hence computation inten-
sive. One of the well-known methods in this category is the graph matching system
[ 15 , 16 ], which is based on dynamic link architecture. Another interesting approach
is the hidden markov model (HMM) [ 17 ] and pseudo 2DHMM [ 18 ]. Blanz et al. [ 19 ]
proposed a method based on 3D morphable model that encodes shape and texture in
terms of model parameters. These parameters are used for face recognition.
7.1.1.2 Appearance Based or Holistic Approach
It is based on statistical approaches, where features are extracted from the whole
image [ 6 , 20 - 22 ]. The extracted feature are from global information, therefore they
are affected from irrelevant information such as hair, shoulders or background which
may affect the recognition results. Yet, the face may still be recognizable by view-
ing only the significant facial region [ 21 ]. In 1990, Kirby and Sirovich [ 23 ]have
developed a technique for efficient representation of images using principal compo-
nent analysis (PCA). Turk et al. [ 24 ] later used this technique for face recognition
using eigenfaces and euclidean distance. Several successive method like eigenphases
[ 25 ], appearance of local regions [ 26 ], Gabor-based PCA [ 27 ], linear or nonlin-
ear discriminant analysis [ 28 - 30 ] use eigenspace transformation based on PCA.
PCA transforms a number of possibly correlated variables into a smaller number of
uncorrelated components. Recently, some applications of ICA [ 31 - 34 ] have been
exploited in image processing and computer vision. It generalize the techniques
of PCA, and proven to be more effective and superior to PCA in many applica-
tions. It defines the independent components from their linear mixture [ 35 - 37 ]of
independent source signals. An important version of ICA, derived from the princi-
ple of optimal information transfer through sigmoid neuron has been used in [ 38 ].