Information Technology Reference
In-Depth Information
early feature-based approaches were based on pure geometry methods, which match
on measurements between features, such as the distance between the eyes or from
the eyes to the mouth. A more sophisticated approach is Elastic Bunch Graph-
Matching [32], which represents faces using labelled graphs, based on a Gabor
wavelet transform. Image graphs of faces are compared with a similarity function.
The technique was shown to be robust to small changes in rotation (up to 22 ). One
advantage of this method is that it requires only a single photo to create the elastic
bunch graph.
Some recent developments have attempted to synthesise appearance-based and
feature-based methods in a hybrid approach. [36] uses semantic features (the eyes
and mouth) to define a triangular facial region. Tensor subspace analysis is applied
only to this region to create the feature vector. This has the advantage of combining
the geometric information from the spatial locations of the local features and the
appearance information from the texture of the facial region. The computational
requirements are much less than Eigenfaces or Tensorfaces alone, and the accuracy
is higher. However, the problems of variations in illumination, pose and scale are
not considered.
1.2.3
Face Recognition from a Video Stream
The machine recognition techniques considered so far were considered in the con-
text of recognising faces from still images. For a surveillance application, we must
also consider the challenges (and opportunities) of recognising faces from a video
stream. In a surveillance video, we can expect that faces will be captured under vari-
able scale, pose and illumination, and that occlusions will be likely. In our proposed
scenario (a Secure Corridor in an airport), we can assume that the environmental
lighting is constant, but there will still be variation in terms of lighting direction and
shadow.
Almost all of the techniques considered so far assume that it is possible to capture
a full-frontal view of the face, but we cannot make this assumption in a surveillance
application. Pose angle can significantly decrease recognition accuracy. [13] tackles
this problem through a pose-estimation step. Faces are extracted from video frames
using an Adaboost classifier, then the same technique is used to locate local features
(eyes) within the face area. The location of the centre-point between the eyes relative
to the centre-point of the detected face area is used to assign a Gaussian weight to
each captured face frame. Face frames which are closer to normal (frontal) pose will
be given a higher weight during the recognition step.
The problem of occlusions is considered in [17]. Part of the face under consider-
ation may be occluded, for example by long hair, sunglasses, a scarf or dark shadow.
The approach taken is “recognition by parts”, where the facial image is broken up
into sub-blocks which can be considered individually. Features based on wavelet co-
efficients are extracted from each sub-block. It is assumed that some of the features
are corrupt, but it is not known in advance which ones. A Posterior Union Model is
used to find the set of features which maximise the probability of a correct match.
 
Search WWH ::




Custom Search