08-08-2012, 03:17 PM
Human face recognition based on multi-features using neural networks committee
[attachment=30610]
Introduction
It is well known that face recognition is a
complex and difficult problem, but it is very
important for surveillance and human–computer
intelligent interaction, security, and telecommunications,
etc. (Wechsler et al., 1998). To solve this
problem, many approaches have been proposed,
e.g., eigen-faces, geometrical features, template
matching, graph matching as well as neural network
and probabilistic method, etc. (Turk and
Petland, 1991; Brunelli and Poggio, 1993; Pessoa
and Leitao, 1999).
Face feature domains
A feature extraction technique referred to as the
interest operator (Wang et al., 1998) is used to find
the directional variances in the horizontal, vertical,
and both diagonal directions for each block in
target images. These directional variances show
the local activity in a block. After obtaining these
directional variance features, we present them to a
neural-network-based classifier.
Conclusions
This paper proposed a new face recognition
method based on multiple feature domains using
neural networks committee machines. The experiments
about the multiple feature domains method
were compared with the single feature domain
methods. And the experimental results for recognizing
the AT&T laboratories database of faces
showed that the face recognition accuracy of the
multiple feature domains method is higher than
that of any one single feature domain. In particular,
when the original images are divided into
blocks, the highest classification accuracy can be
attained. Obviously, our proposed method indeed
improves the classification accuracy.
[attachment=30610]
Introduction
It is well known that face recognition is a
complex and difficult problem, but it is very
important for surveillance and human–computer
intelligent interaction, security, and telecommunications,
etc. (Wechsler et al., 1998). To solve this
problem, many approaches have been proposed,
e.g., eigen-faces, geometrical features, template
matching, graph matching as well as neural network
and probabilistic method, etc. (Turk and
Petland, 1991; Brunelli and Poggio, 1993; Pessoa
and Leitao, 1999).
Face feature domains
A feature extraction technique referred to as the
interest operator (Wang et al., 1998) is used to find
the directional variances in the horizontal, vertical,
and both diagonal directions for each block in
target images. These directional variances show
the local activity in a block. After obtaining these
directional variance features, we present them to a
neural-network-based classifier.
Conclusions
This paper proposed a new face recognition
method based on multiple feature domains using
neural networks committee machines. The experiments
about the multiple feature domains method
were compared with the single feature domain
methods. And the experimental results for recognizing
the AT&T laboratories database of faces
showed that the face recognition accuracy of the
multiple feature domains method is higher than
that of any one single feature domain. In particular,
when the original images are divided into
blocks, the highest classification accuracy can be
attained. Obviously, our proposed method indeed
improves the classification accuracy.