文档介绍：F ace Recognition using a Hybrid Sup ervised
Unsup ervised Neural
Net w ork
Nathan In trator Daniel Reisfeld Y ehezk el Y esh urun
Departmen t puter Science
T el
Aviv Univ ersit y
Ramat Aviv
Israel
Revised
June
Abstract
A system for automatic face recognition is presen ted
It consists of sev eral steps
Automatic
detection of the ey es and mouth is follo w ed b y a spatial normalization of the images
The classi
cation of the normalized images is carried out b y a h ybrid
sup ervised and unsup ervised
Neural
Net w ork
Tw o metho ds for reducing the o v er
tting
mon problem in high dimensional
classi
cation sc hemes
are presen ted
and the sup eriorit y of bination is demonstrated
Key w ords
F ace recognition
w orks
In terest p oin ts
Symmetry op erator
T o app ear
P attern Recognition Letters
In tro duction
Automatic face recognition has gained m uc h atten tion in recen t y ears
due to the v ariet y of p oten tial
applications
and the increase putational p o w er whic h enables e
ectiv e implemen tation of
algorithms
T raditionally
face recognition w as based on extracting certain features
e
g
spatial lo cation of
facial features and their geometrical relations
These features are detected either man ually
or b y automatic algorithms
Another approac h
is based on direct pro cessing
of the grey lev el images
A review of face pro cessing systems could b e found in
The task of recognizing faces is inheren tly a classi
cation problem in high dimensional feature
space
and th us sub ject to the
curse of dimensionalit y
whic h essen tially sa ys that the n um b er
of training patterns needed for robust classi
cation
should b e restrictiv ely high
Regarding an image merely as a matrix and lo oking for algebraic in v arian ts
reduces the
dimensionalit y
Ho w ev er
suc h algebraic constrain ts can b e designe