Professional Documents
Culture Documents
Face Recognition
Face Recognition
1 (2002) 47–52
c World Scientific Publishing Company
TAKUMA TAKEHARA
Department of Psychology, Doshisha University
Kyoto 602-8580, Japan
takehara@psychology.doshisha.ac.jp
FUMIO OCHIAI
Department of Human and Cultural Studies
Tezukayama University, Nara 631-8501, Japan
NAOTO SUZUKI
Department of Psychology, Doshisha University
Kyoto 602-8580, Japan
Abstract
Following the Mandelbrot’s theory of fractals, many shapes and phenomena in nature have been
suggested to be fractal. Even animal behavior and human physiological responses can also be
represented as fractal. Here, we show the evidence that it is possible to apply the concept of
fractals even to the facial expression recognition, which is one of the most important parts of
human recognition. Rating data derived from judging morphed facial images were represented
in the two-dimensional psychological space by multidimensional scaling of four different scales.
The resultant perimeter of the structure of the emotion circumplex was fluctuated and was
judged to have a fractal dimension of 1.18. The smaller the unit of measurement, the longer
the length of the perimeter of the circumplex. In this study, we provide interdisciplinarily
important evidence of fractality through its application to facial expression recognition.
47
48 T. Takehara et al.
Prototype Prototype
1. INTRODUCTION face face
Surprise
Fear Excitement
Anger Happiness
Sadness Calm
Sleepiness
(a)
Happiness Calm
(b)
Fig. 2 (a) Schematic representation of eight prototype facial expressions on the perimeter of a circumplex. Note that the
distances between prototypes are not exactly the same. (b) An example of Happiness-Calm pair of morphing process and its
morphed images in condition (i). Physical increment is kept constant. The other pairs were carried out in the same way.
condition (i), 19 for (ii), 31 for (iii), and 28 for (iv). adjacent prototype face pairs positioned on the
They volunteered to be research participants to get perimeter in each of the four different scales using a
a credit for a psychology course. All had normal or morphing software (Morpher for WindowsTM ). In
corrected-to-normal vision. condition (i), each interval between adjacent pro-
totypes was divided into four segments and three
morphed facial images were generated. In total,
2.3 Stimuli
32 facial images (eight prototypes and 24 morphs)
Based on the arrangement illustrated in Fig. 2, were generated around a circumplex for condi-
morphed facial images were generated between all tion (i). Similar to this condition, morphed facial
50 T. Takehara et al.
1.5
images were generated for the other three con- Excitement
Surprise
ditions: 40 facial images (eight prototypes and
32 morphs) for condition (ii), 48 facial images 1.0
(eight prototypes and 40 morphs) for (iii), and 64 Fear
facial images (eight prototypes and 56 morphs) for 0.5
(iv). The particular advantage of the morphing is
that it is possible to create highly objective facial Happiness
images and easy to control the amount of physi- 0.0
2.4 Procedure
-1.5
For each of the four experimental conditions, Sleepiness
participants were asked to rate all facial images on -2.0
a six-point Likert-type scale (from “not at all” to -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
“very so much”) about each item of the eight pro-
totype emotions of Happiness, Calm, Sleepiness, Fig. 3 Configuration of prototype and morphed facial
images in the two-dimensional psychological space. This ex-
Sadness, Anger, Fear, Surprise and Excitement.
ample illustrates a resultant configuration of condition (ii).
In rating, facial stimuli were presented one at a Black rectangles indicate the prototypes and white rectan-
time in a random order on a 14.1-inch LCD with gles morphs. Adjacent faces are tied with a straight line.
240×320 pixels. The distance between a participant All morphs are arranged between their prototypes, but the
and a display was about 50 cm. Data were collected perimeter of a circumplex seems to be complex and wiggled
rather than smooth. The other conditions showed the same
by a Compaq personal computer (DESKPRO;
tendency.
Pentium II Processor 350 MHz) with a two-button
mouse and by a rating software made by an Inprise
C++ compiler (Borland C++ Builder Version 4).
No feedback was given. prototypes fell in a roughly circular order and were
rated in the same order under all conditions,3,4,6,7,22
and all morphs were plotted between their parent
3. RESULTS prototypes corresponding to their morphing pro-
portions. The resultant configuration in condition
In order to visualize the assumed two-dimensional (ii) is shown in Fig. 3. Roughly speaking, each
psychological space framework, we need to spec- configuration may appear to form a circle, but no-
ify the relative positions of faces within the psy- tice that the perimeter is complex and somewhat
chological space. This can be accomplished with wiggled rather than smooth. The distance of the
multidimensional scaling (MDS), which derives a perimeter of a circumplex is defined here as the to-
spatial representation of the stimuli from psycholog- tal distance, which is the sum of distances between
ical data, such as ratings of perceived similarity.21 adjacent facial images.
For each of the four experimental conditions, rating If the distance of the perimeter were exactly the
data were used to calculate a matrix of dissimilari- same for the four conditions despite the different
ties among eight facial expression categories, which scales, it would not have fractal properties, suggest-
was then subjected to MDS to find coordinates of ing that the perimeter has a fixed and characteristic
all facial images in two-dimensional space. For the length scale like a complete circle. On the contrary,
two-dimensional configuration, all RSQ values were if it were revealed that as the scale was made smaller
sufficiently high: 0.978 for condition (i), 0.971 for and smaller, the distance of the perimeter tended to
(ii), 0.976 for (iii), and 0.977 for (iv). RSQ is a mea- increase steadily without bounds, we would obtain
sure of the proportion of the variance in the data a fractal dimension, indicating that the perimeter
accounted for by the MDS solution. Similar to the had no characteristic length scale, and was fluc-
results of many previous studies, the positions of the tuated by a fractal dimension. To test this, we
Facial Expression Recognition 51