their facial expressions. 
used in a learning process. 
4.2 Learning Self-organizing Maps 
those used in a learning process. 
rectangular segments is the same as the size of those 
The size of the 
The learning time self-organizing maps require is 20000. Table 1 Details of Images Used for Learning 
  
Table 1 shows detailed conditions of learning. 
  
Rectangular segments corresponding to each facial 
  
  
segment are selected from facial images. Table 2 shows 
structure of self-organizing maps. The self-organizing 
  
  
maps are designed 1-dementional. It makes a following 
analysis simple. 
  
  
  
  
4.3 Feature Extraction of Facial Expressions 
  
Features are extracted by using self-organizing maps Size of 
  
learned, in accordance with the conditions discussed 
  
above. Rectangular segments for each facial segment 
  
  
  
Size of Facial Image 320X240 [pixel] 
Position of Right Eyebrow 106, 51 [pixel] 
Size of Right Eyebrow 60X 35 [pixel] 
Position of Right Eye 115, 81 [pixel] 
Size of Right Eye 45X 25 [pixel] 
Position of Mouth 130,141 [pixel] 
Size of Mouth 80X 40 [pixel] 
Table 2 Structure of Self-organizing Map 
Size of 
Input Neuron Output Neuron 
Map of Eyebrow | 60X35 [pixel] | 40X 1 [pixel] 
Map of Eye 45 X 25 [pixel] | 40X 1 [pixel] 
  
  
are selected from input images at the same positions as 
Map ofthe eyebrow Neuron number 
    
  
  
Map of the eye Neuron number 
sg A OL SN PO HAN HAM CR i6 
V MANU ARN WOE A Tl Sh non 
Sth vl nt RAN A onl Dll 21200 
i 
Map of the mouth 
ENENNENN (o 
IES 11 ~ 20 
21~30 
31~40 
Figure 4 Subjects and Their Facial Expressions 
  
Victory neuron 
number of 
eyebrow 
Hu SE PP Se RTS ss 
Ot eyebrow.. bile cd zi 7 br SE ati 
Self-organizing map.L c7 3-1. --32---L--1--] | 
  
  
  
  
Self-organizing map | 
of eye number of eye 
Victory neuron 
number of mouth 
Time 
| Victory neuron 
  
  
Self-organizing map : 
of mouth 
  
  
  
  
  
(a) The movements of the victory neuron for each facial segment 
S - 
a 9 
3 < 3 10 
o 3 
9 = os. 
m o c 20 
Sc co 
S vs 
= S 30 
2 9 
40 
0 
L Ls 
Cf, Cte, 
  
(b) The movements in 3-dementional space 
Figure 5 The movements of the Victory Neuron(Anger) 
446 
Neuron number 
  
(c) The movements when strongly expressed 
  
[Map of 1 
Figure 4 
recognition 
neuron wh 
input. Th 
shapes for 
of the ea 
movement: 
Figure (c) 
while the fe 
Figure 6 s 
different kir 
of the victo 
expression 
6.1 Validit 
The facial 
ware base 
express. 
accurately 
questionne 
to watch a 
expression 
findings. 
interested 
recognitior