See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/10582051

Textural properties corresponding to visual perception based on the

correlation mechanism in the visual system

Article  in  Psychological Research · September 2003

DOI: 10.1007/s00426-002-0113-6 · Source: PubMed

CITATIONS READS

35 503

3 authors, including:

Kenji Fujii Yoichi Ando

11 PUBLICATIONS   196 CITATIONS   
Graduate School of Kobe University
329 PUBLICATIONS   2,970 CITATIONS   
SEE PROFILE
SEE PROFILE

Some of the authors of this publication are also working on these related projects:

DSSF3 Diagnostic system for sound fields View project

Neurally used theory for speech percepts and applications (tentative book title) View project

All content following this page was uploaded by Kenji Fujii on 03 June 2014.

The user has requested enhancement of the downloaded file.

Psychological Research (2003) 67: 197–208
DOI 10.1007/s00426-002-0113-6
O R I GI N A L A R T IC L E
Kenji Fujii Æ Shinofu Sugi Æ Yoichi Ando
Textural properties corresponding to visual perception based
on the correlation mechanism in the visual system
Received: 2 May 2001 / Accepted: 22 July 2002 / Published online: 5 March 2003

Springer-Verlag 2003
Abstract We present a set of texture parameters that
correspond to perceptual properties of visual texture.
For machine vision or a computer interface, it is
important that the computational measurements of
texture correspond well to the perceptual properties. To
understand the mechanism of our visual system, it is
important to know how we extract or characterize
information for texture perception. In this study, we
show that the autocorrelation function (ACF) analysis
provides useful measures for representing three salient
perceptual properties of texture: contrast, coarseness,
and regularity. The validity of the ACF analysis was
examined by comparing the calculated factors to the
subjective scores collected for various kinds of natural
textures. The effectiveness of the analysis depends on the
structure of the estimated ACF. When a texture has a
harmonic structure, the estimated ACF has periodical
peaks corresponding to the periods of the texture. Both
perceived coarseness and regularity are strongly related
to these peaks in the ACF. However, the estimated ACF
does not have a periodical structure when the texture is
random. In this case, the texture coarseness and regu-
larity are represented by the decay rate of the ACF.
Introduction
In this paper, we present a set of texture parameters that
correspond to perceptual properties of visual texture.
The purpose of this study is to bridge the gap between
K. Fujii (&) Æ S. Sugi Æ Y. Ando
Ando Laboratory, Graduate School of Science and Technology,
Kobe University, Rokkodai, Nada, Kobe, 657–8501 Japan
E-mail: fujii@ymec.com
Tel.: +81-3-33705160
Fax: +81-3-33706348
K. Fujii
Now at Yoshimasa Electronic Inc.,
Daiichi-Nishiwaki Building, 1-58-10 Yoyogi,
Shibuya, Tokyo, 151-0053 Japan
the research on computational texture modeling with
texture perception and human visual system. The un-
derlying problem is how the human visual system
characterizes salient texture properties. Several previous
studies have focused on the texture properties either
through the study of perception or through computa-
tional models. But the relationship between perceptual
and computational properties remains unclear. To
understand the mechanism of our visual system, it is
important to know how we extract and characterize the
information to perceive textural properties. For many
computer vision applications, perceptual consistency is
important. For example, it is necessary that computa-
tional texture properties correspond well to the percep-
tual properties in an image understanding system or a
context-based image retrieval system. Therefore, we
believe that this study contributes both to an under-
standing of the operation of human visual system, as
well as to computer vision.
Previous research
Historically, textural features were computed by first-
order and second-order statistics of the image, such as
average and variance, or a co-occurrence matrix
(Haralick, 1979). These approaches were mainly driven
by the general texture model proposed by Julesz (1962).
Another approach was to describe texture as primitive
and to analyze it in terms of its placement rules. The
main deficiency of the structural methods is that they are
incapable of capturing or generating the randomness
that natural textures almost possess. Since the 1980s, the
study of texture models has been based on the random
field statistics (Cross & Jain, 1983; Mao & Jain, 1992)
such as Markov random field or autoregression models.
These models can describe texture by using a small
number of parameters. But these parameters are not
based on the mechanism of the human visual system,
and therefore they can hardly represent multiple textural
properties salient to the visual perception.
198
More recent texture models are motivated by the
psychophysical and physiological findings on the early
stage of human visual system. These texture models are
based on the psychophysical evidence that two textures
are often difficult to discriminate when they produce a
similar distribution of responses in a bank of orientation
and spatial-frequency selective linear filters (Turner,
1986; Malik & Perona, 1990). In most of the articles (e.g.
Heeger & Bergen, 1995; Zhu, Wu & Mumford, 1998),
texture is parameterized by the first-order statistics (e.g.
histograms) of a set of an appropriately chosen multi-
scale and multi-orientation filter outputs. The applica-
tion of these sets of parameters to texture synthesis has
been successful in generating a new texture that matches
the appearance of a given sample. The main disadvan-
tage of this method is the difficulty in synthesizing pe-
riodic or structured textures. Another problem that
remains to be solved is how to select a set of features to
best characterize the images and how to integrate the
output of many filters for further processing.
Considering the effective computational strategy in
our visual system, it is quite possible that not all the
information carried out by the high-dimensional sensory
representation is preserved for rapid judgments of nat-
ural textures. Furthermore, when similarity of more
than two patterns is under consideration, the task of the
computational model is not only to determine whether
the patterns are different, but also to specify how much
they differ. Therefore, after the initial linear filtering
stage, a more sophisticated process of information
characterization and quantification is most likely in-
volved. A number of studies used higher order percep-
tual features to explain the human judgments of the
similarities in the appearances of textures. Implied in
these studies is a suggestion that not only are perceptual
features useful in judgments of complex textures, but
that the number of useful features is small.
For example, Tamura, Mori, and Yamawaki (1978)
measured human discrimination performance with tex-
ture features and compared these measures with com-
putational results. From the description in the literature
and from observations of the Brodatz texture album
(Brodatz, 1956), six features were chosen that corre-
spond to visual perception: contrast, coarseness, regu-
larity, roughness, directionality, and line-likeness. The
emphasis was on developing computational analogues of
these attributes, but their approach had drawbacks.
They only modified already developed computational
features and combined several features to have a close
relationship to a specific property. As a result, this
approach failed to clarify how visual properties are
related to particular physical properties. Amadasun and
King (1989) carried out a similar study, but their com-
putational features had similar drawbacks. Rao and
Lohse (1996) developed a classification method for
perceptual texture. Based on psychological similarity
judgments, they constructed a three-dimensional space
for texture classification. The three orthogonal dimen-
sions were repetitive vs. non-repetitive; high-contrast
and non-directional vs. low-contrast and directional;
and granular, coarse, and low-complexity vs. non-
granular, fine, and high-complexity. Similar experiments
conducted by Cho et al. (2000) suggested that the di-
mensionality of perceptual texture space was at least
four. They described four orthogonal attributes, namely,
coarseness, regularity, contrast, and lightness.
The studies reviewed suggested that the important
properties of perceptual texture could be described by
using three or four independent factors. Among these
orthogonal dimensions, the importance of perceived
regularity or periodicity is emphasized in most of the
studies. Hence, it is desirable for a texture modeling to
describe regularity. Actually, there are several studies
emphasizing the importance of periodical structure in
texture modeling and synthesis. Francos, Meiri & Porat
(1993) decomposed texture spectrum into harmonic
(periodical), evanescent (directional), and indeterministic
(random) components for texture synthesis. Liu & Picard
(1996) extended a similar method for an image retrieval
system. Their methods treat periodical and random
components separately and succeeded in representing
many kinds of natural textures. Portilla & Simoncelli
(2000) included local autocorrelation in their synthesis
parameter to represent periodic structures in a texture.
Also in vision research, the role of periodical structure
has been examined. Uttal (1975) emphasized the impor-
tance of periodical structure in form perception. He
conducted an experiment of form detection with dot
patterns, measuring the detectability of dot patterns
masked by noise patterns. He found that the regularity of
dots in a pattern increased the detectability of the pat-
tern. The autocorrelation principle was applied to explain
this result. In his autocorrelation model, the periodicity
of the signal is emphasized and the random noise is
minimized. As a result, only the periodical signal is
extracted and then perceived as a form. Recently, Ben-Av
and Sagi (1995) presented an autocorrelation function
(ACF) model for perceptual grouping. They used matrix
patterns of discrete elements to quantify the grouping
law. The elements were arranged in arrays giving rise to
two possible perceptual organizations, either horizontal
rows or vertical columns. In the model, the ACF was
calculated in vertical and horizontal directions. Percep-
tual grouping occurred in the direction in which the
higher degree of correlation appeared. Their results imply
that there is a correlation mechanism involved in the
process of perceptual grouping. They proposed a possi-
bility that orientation filters (e.g. Gabor filters) with long-
range interactions between them could be used for the
estimations of directional autocorrelations.
Purpose of the study
In this study, we tried to develop a set of measures for
texture properties that correspond to visual perception.
We measured psychophysical scales for the textural
properties of contrast, coarseness, and regularity. These

three properties were chosen because they were found
important in the studies of texture similarity judgments.
In particular, motivated by the discussion described
above, we considered that the periodical structure and
the correlation structure are important for visual per-
ception of texture. Therefore we analyzed natural tex-
tures by using the ACF and extracted several measures
to capture the periodical structure. The ACF may not be
a novel and obvious method for texture analysis, but we
show that the ACF analysis provides useful measures for
representing the three salient perceptual properties. The
validity of the ACF analysis is examined by comparing
the calculated factors with the subjective scores collected
for various kinds of natural textures.
Autocorrelation analysis of natural textures
Definition of ACF
The ACF of a two-dimensional texture pattern is defined
as
XX
M
1 N
1
UðDx; DyÞ ¼ pðx; yÞpðx þ Dx; y þ DyÞ=MN ð1Þ
x¼0 y¼0
where p(x, y) is the input signal and p(x+Dx, y+Dy) is
the shifted version of the input. The analyzed image is
zero-meaned before the calculation to remove the DC
components. ‘‘M’’ and ‘‘N’’ refer to the signal size in
horizontal and vertical directions. In this study, M = N
= 256. Equation (1) is the correlation of the signal itself.
For computational efficiency, the ACF is computed as
an inverse FFT of the image power spectrum. Usually,
the ACF is normalized as
! (4) de: the effective range of the ACF (defined as the
X NX
M
1
1
2
/ðDx; DyÞ ¼ UðDx; DyÞ= pðx; yÞ =MN ð2Þ
x¼0 y¼0
The denominator in equation (2) is the maximum value
of the ACF and represents power of the image. After
this normalization, the ACF takes the maximum value
of 1.0 at the origin.
Because the ACF is mathematically equivalent to the
power density spectrum, we can generally derive the
same information from both the ACF and power spec-
trum in space and frequency domain respectively. For
example, a vertical stripe pattern has its fundamental
Fourier components along the horizontal axes. The
ACF of the pattern has periodicities that correspond to
the reciprocal of its fundamental frequency. However,
there is a certain case in which the ACF differs greatly
from the spectrum. When the fundamental component is
subtracted from the stripe pattern (missing fundamental:
MF), we can still see the period of the fundamental
component (Henning et al., 1975). The frequency anal-
ysis fails to explain this percept because the fundamental
component is absent. Instead, the ACF of this MF
pattern is found to maintain a period of the fundamental
199
component. The perceptual mechanism of the MF pat-
tern is still under discussion (Badcock & Derrington,
1989; Hammett & Smith, 1994), but it implies that the
ACF is effective to detect the spatial periodicity of pat-
terns. The ACF alone is insufficient to capture the
structure of many natural textures because the ACF
does not contain the local phase information. Never-
theless, for the purpose of quantification of salient per-
ceptual properties, the ACF is still useful to represent
periodic structure and correlation structure within the
image.
Factors extracted from the ACF
From the calculated ACF, we extracted four typical
features. Here we only considered the one quadrant
because the ACF is symmetrical. As shown in Figure 1,
the ACF decays from the origin to the outwards. To
simplify the calculation, we only considered two one-
dimensional ACFs along the x and y axes from the
origin. We assumed that the properties of the two-
dimensional ACF are held in the one-dimensional
ACFs. This is because most natural texture is considered
to be isotropic, in which all orientations occur with the
same probability. Consequently, the ACF could be
assumed to be circular symmetric. Even for the aniso-
tropic texture, at least one of our directional ACFs can
hold the periodical structure. Below, we define the four
extracted factors and then describe them in more detail.
(1) F(0): the autocorrelation at shift value Dx = Dy = 0.
(2) d1: the displacement of the maximum peak in the
ACF except for the origin.
(3) /1: the amplitude of the maximum peak.
displacement at which the envelope of the nor-
malized ACF decayed below 0.1).
Only the factor F(0) was calculated from the two-di-
mensional ACF (i.e. denominator in equation 2), and it
is shown on the dB scale (10 log10F(0)). The other three
factors were calculated by the algorithm as follows.
First, directional ACFs along the x and y axes are ex-
tracted from the two-dimensional ACF. Then the max-
imum peak values (except the origin) of both are
calculated and compared so as to choose the more pe-
riodical one. From the chosen ACF, the displacement
and the amplitude of the maximum peak are picked out
as d1 and /1 (see Fig. 1b). The value of de is determined
from the cross point of the envelope at 0.1 as shown in
Fig. 1b. This is done by exponential fit to the ACF peaks
above 0.1. When the given ACF has no peaks above 0.1,
only the initial decay range was fitted.
The value of F(0) corresponds to the root-mean-
square (RMS) of the signal, because the ACF was cal-
culated from the zero-meaned image. Thus, it is assumed
that F(0) corresponds to the perceived contrast. The
displacement and the amplitude of the maximum peak
200
Fig. 1 A calculated ACF and
factor definitions. (a) Two-di-
mensional ACF of texture D3
in Fig. 2. (b) and (c) One-
dimensional ACF for the x and
y directions. By comparing the
heights of the maximum peaks
in both directions, we extracted
three factors from the ACF for
the x direction
(d1 and /1) are related to the periodicity of the image.
The value of d1 is a reciprocal of the fundamental fre-
quency, and /1 is the strength of the harmonic compo-
nents. A large value of d1 means that low spatial
frequency is perceived, and also large values of /1 mean
that the image is more periodic. Therefore we assumed
that the strength of coarseness and regularity could be
represented by d1 and /1 respectively. Because the
effective range of the ACF is related to the height of the
peaks, de might also be related to the perceived regu-
larity. The values of d1 and de were normalized based on
the image size. The visual angle (degree) of the values is
calculated by applying image width (cm) and dividing by
viewing distance (cm).
Evaluation of ACF analysis
For evaluation of the proposed analysis, we analyzed 16
textures from the ‘‘Brodatz texture database’’ (Brodatz,
1966) that were used by Tamura et al. (1978). As men-
tioned before, Tamura et al. measured perceptual scales
for six textural features including contrast, coarseness,
and regularity, which are the features we are interested
in. To examine the validity of the ACF analysis, the
scaling of texture features measured by the ACF factors
were compared with the psychophysical scaling. Then,
the subjective scores for the three texture properties of
contrast, coarseness, and regularity were correlated with
the ACF factors. Our hypothesis is that perceived con-
trast would correlate with the ACF measure of F(0),
coarseness with d1, and regularity with /1.
Data collection
In Tamura et al. (1978), psychophysical experiments
were conducted to establish the ordering of 16 texture
samples based on the six properties. In the experiment,
48 human subjects gave their paired comparison judg-
ment for all possible pairs of 16 texture samples. Sub-
jects were required to make decisions according to the
six properties, i.e., to choose the coarser, the higher
contrast, the more regular, the more directional, the
more line-like, and the rougher sample of each pair.
Note that these six properties have both extremes in
their concept, e.g., coarse versus fine for coarseness, to
give well-distributed results. The collected data were
processed by applying ‘‘the law of comparative judg-
ment’’ (case V; Thurstone, 1927) for constructing the
one-dimensional psychometric scale value (SV).
They showed their subjects’ average scores as scales
on one-dimensional distributions in their article. We
obtained them from the article by using an image
scanner, and re-scaled linearly between 0 and 1. Ana-
lyzed texture set includes a variety of natural and man-
made objects (Figure 2). The analyzed data consists of
256 · 256 pixel images with 8-bit (256) gray levels. The

Fig. 2 Texture set used by
Tamura et al., (1978)
ACF was calculated and four factors were extracted for
each image.
Results
We compared the obtained data with our computed
results. The subjective scores (scale value: SV) for the
contrast, coarseness, and regularity were correlated with
the factors extracted from the ACF: F(0), d1, and /1.
Figure 3 shows the relationship between subjective
scores for each property and the ACF measures. The
ACF factors are shown on a logarithmic scale because of
its goodness for linear regression. Each point represents
the SV of each sample. Straight lines represent the linear
regressions. We can see that the scores of the ACF
factors are in good agreement with the subjective scores.
In particular, the correlation coefficient was very high
(r = .89, p <0.01) for perceived contrast and F(0). As
described above, the ACF factor F(0) corresponds to the
image power. This is consistent with the previous studies
that found that textures with the same RMS are per-
ceived as the same contrast (Mayhew & Frisby, 1978;
Tiippana & Näsänen, 1999).
201
The correlation between the SV of regularity and the
value of /1 was also high (r = 0.86, p <0.01). But
several samples with values of log (/1) below –1 (left side
in Fig. 3b) showed relatively large variance from
regression line. The correlation only in this range was
very low (r = 0.35). This implies that regular-random
dimension can hardly be represented by the value of /1
alone, especially for random texture. Perceived regular-
ity was also related to the value of de, because of the high
correlation between /1 and de as described in the defi-
nition of the ACF factors.
The correlation between perceived coarseness and d1
was 0.70 (p <0.01), which is worse than the previous
two results. There was also some discrepancy for sam-
ples in which /1 had a low value. This is perhaps because
d1 is strongly dependent on the value of /1. When the
ACF contains high peaks due to a high regularity of
the image, /1 and d1 can be easily determined and the
resultant value of d1 accurately represents the dominant
period of the image. But when the image is random, the
ACF decays immediately with no particular peaks.
Consequently, the determined value of d1 might not
reflect the period of the image. This problem is discussed
in detail later.
202
Fig. 3 Relationships between the scale values (SV) of textural
properties and corresponding ACF factors. Each dot represents a
scale value for each texture, and the straight lines express linear
regression
Multiple regression analysis
It was found that the ACF factors of F(0), d1, and /1
could represent the perceptual properties of contrast,
coarseness, and regularity. But there remains some vari-
ance that cannot be explained by single factors alone. A
portion of this unexplained variance might be due to the
properties not being truly one-dimensional. Actually,
high correlation was found (r = 0.66) between the per-
ceived coarseness and contrast. We used multiple
regression analysis to examine whether the plural factors
affect the perceptual properties. The aim of the multiple
regression analysis is to predict the behavior of a depen-
dent variable by using a weighted sum of independent
Fig. 4 The relationships between predicted SVs by means of
multiple regression analysis and measured SVs of contrast and
coarseness
predictor variables. In our case, we tried to predict the
subjective scores by using a linear combination of four
ACF factors. To obtain the optimal model for any given
number of predictors, all 11 possible regression equations
were examined. The correlation coefficients and signifi-
cance levels were used to determine the goodness of fit.
For perceived contrast and coarseness, the best pre-
dictors were F(0) and d1. The other factors also had
some use, but their contribution was not significant. As
shown in Figure 4, the higher correlation coefficients
were obtained (r =.96, p <0.01 for contrast, and r =.83,
p <0.01 for coarseness). The partial regression coeffi-
cients of F(0) and d1 were .98 and .36 for the contrast,
.50 and .79 for the coarseness. It is considered that the
perceived contrast and coarseness are represented
mainly by F(0) and d1. This result is consistent with the
result that perceived contrast is relatively high for and
coarser texture. For regularity, only a single factor /1
was significant (p <0.01) in all combinations. Even
though the other factors also had some contribution, the
resultant correlation gained little by the addition of these
factors.
Discussion
In this section, the proposed ACF measures were com-
pared with psychophysical measurement of three texture

properties. The results show that the ACF factors can be
used to represent the perceptual properties of contrast,
coarseness, and regularity. The correlation between
human and computational scores showed similar or
higher levels compared with those found in Tamura
et al’s original results. It is surprising that these simple
measures account so well for the subjective attributes of
complex natural textures. The remaining problem is,
however, how to improve the correspondence between
the ACF measure and the perceived regularity and
coarseness for random texture.
As described above, the ACF for random texture
does not have a periodical property. Consequently, it is
difficult to extract information related to regularity and
coarseness from only the maximum peak of the ACF.
This is because we cannot guarantee that the maximum
peak is periodic. In this case, the value of de (the
effective range of the ACF, defined as a ten-percent
decay rate of the ACF) might be a parameter to rep-
resent the perceived regularity and coarseness. If an
image does not contain any harmonic components, the
value of de corresponds to the slope of its spectrum as
shown in Figure 5. It is known that the power spec-
trum of natural images can be represented as a func-
tion of spatial frequency f and its slope b on the log-log
scale (i.e. P  f )b). When the spectrum is flat (white
noise), the value of de becomes very small (theoreti-
cally, de fi 0). When the spectrum has a slope (corre-
lated noise), the value of de becomes larger in
proportion to the value of the slope. It is quite possible
Fig. 5 Example of correlated
noises (256 · 256 pixels) filtered
with different slope (b = 0.5,
1.0, 1.5, and 2.0) of power
spectra (P  f )b) on the log-log
scale. Power spectra and the
ACF are also shown as func-
tions of radial frequency and
radial distance, because of their
circular symmetry
203
that the more the correlation increases within an image,
the more we can perceive the high degree of regularity
for that image. Also, it is possible that de is large in
coarse texture, and small in fine texture. In the next
psychophysical experiment, we used a set of random-
looking textures to examine whether the value of de
could represent the perceived coarseness and regularity.
Psychophysical experiment
The purpose of this psychophysical experiment is
twofold. First, we wanted to examine whether human
subjects could detect small differences in the properties
and whether the results of the ACF analysis corre-
sponded to the subjects’ judgments. Second, we wanted
to examine whether the ACF factors are useful in an-
alyzing textures with less harmonic structures. For
these purposes we used a set of botanical textures with
a high degree of similarity between them and less
harmonic structures. Our hypothesis in this experiment
is that perceived contrast would correlate with the ACF
measure F(0), and coarseness and regularity would
correlate with de.
Methods
Subjects Ten subjects (5 males and 5 females) participated in the
experiment. Their ages ranged from 22 to 26. All had normal or
corrected-to-normal visual acuity. With the exception of the au-
204
Fig. 6 The stimulus set used in
the psychophysical experiment
thors (FK and SS), the rest of the subjects were unaware of the
purpose of the study.
Apparatus and stimulus The images were presented on a CRT
display under normal indoor lighting conditions. For the paired
comparison judgment, each stimulus consisted of horizontally lo-
cated two texture pairs within a dark background. The viewing
distance was approximately 100 cm, which resulted in each texture
image subtending a visual angle of 6 · 6 degree. The sample texture
set used in the experiment consisted of 12 botanical textures as
shown in Figure 6. The pictures were taken by a digital camera and
were transformed into 256 · 256 pixel images with 8-bit gray levels.
All the samples were then analyzed by using the ACF and four
factors were extracted. Compared with Tamura et al.’s texture set,
our samples had a higher degree of similarity. Consequently, the
extracted factors fell into a narrow range (Table 1). Furthermore,
the values of /1 were much lower for our stimuli. This means that
the structure of our sample textures was more random.
Procedure Subjects were given a brief explanation of the basic
concept of the three texture properties (the explanations were given
in Japanese, except for the names of properties such as ‘‘coarseness;
coarse versus fine’’, which were given in English). Each property
was defined as an internal attribute of the image, and the subject’s
task was to compare the strength of each property for a pair of
images. All possible pairs from the 12 images (66 pairs) were pre-
sented to the subjects in a random order in one session. During a
presentation (10 s) they were asked to make their decisions about
the three properties separately and write down their answers on the
check sheet, i.e., to choose the pattern that was coarser, had a
higher contrast, and was more regular. Before the next presentation
began, a 2 s interval was given. It seemed this would be a difficult
task for the subjects because they have to make three different
judgments in a short duration. But the pilot study confirmed that
the subjects made as reliable judgments as in the case of separate
presentations for individual judgment. Following the training ses-
sion, all subjects had eight series of sessions, giving a total of 528
comparisons for each feature.
Data analysis The obtained data were processed by applying ‘‘the
law of comparative judgment’’ (case V; Thurstone, 1927). This law
was used to produce a one-dimensional scale value (SV) for each
stimulus from the total matrix of superiorities collected from the
paired comparisons.
Table 1 ACF factors for textures used in Tamura’s experiment and
ours. Mean values are shown with standard deviation
Sample set F(0) /1 d1 de
Tamura et al. –37.99 (2.78) 0.21 (0.23) 0.11 (0.1) 0.19 (0.36)
Present study –39.08 (1.61) 0.09 (0.05) 0.03 (0.01) 0.02 (0.01)
Table 2 Correlation matrix of the scale values of contrast,
coarseness, and regularity
Property contrast coarseness regularity
contrast –
coarseness 0.672* –
regularity 0.017 –0.363 –
*p <.05
Results The results were reconfirmed by the goodness of fit
(Mosteller, 1951), and the agreement of all subjects’ judgments was
tested by the chi-square test (p <0.05). The correlation coefficients
between the subjective scores for each feature are shown in Table 2.
We found the following:
(1) For all perceptual properties tested, the subjects’ judgments are
reliable and there are certain underlying criteria upon which the
tested subjects agreed. This does not mean that the same data
were obtained for all subjects in the experiment. Individual
differences must be considered.
(2) In the regularity score, a relatively high inter-subject variance
was observed. There might be several reasons for this variance.
First, some of our subjects may have interpreted ‘‘regularity’’ in
a different manner, or the concept of regularity itself might have
been difficult to define for the tested textures. Second, because
there was little difference in material or appearance between the
sample textures, the subjects might have difficulty in comparing
regularity. Thirdly, each subject might have used different cues
for the judgment of regularity. We specified regularity as a

property of the placement rules. But the variations of the ele-
ments reduce the perceived regularity as a whole, especially in
the case of natural textures. As a result, it is assumed that the
subjects perceived a high regularity of those textures for which
we can describe texture elements easily.
(3) There was a high correlation between the SV of contrast and
coarseness (r =.67, p <0.05). This means that coarser texture is
perceived to have a higher contrast. From these observations,
we concluded that our experiment was valid for measuring
human perception of textural properties, even though there
were some individual differences. It was found that subjects
could detect small differences in the texture properties reliably.
Next, we compared the psychophysical data and computed
texture factors.
Figure 7 shows the relationship between the SV for each property
and the ACF measures. As shown in Fig. 7a, there was a high
correlation (r =.81, p <0.05) between the SV of contrast and F(0),
consistent with the results of the previous experiment. The RMS
power could be a useful parameter to represent perceived contrast
even for the set of natural texture restricted in the narrow range. In
Fig. 7b, the result for coarseness is also related to de. There is a
moderate correlation, allowing some variance (r = .63, p = 0.053).
The perceived coarseness was also related to the value of d1, but the
correlation between d1 and SV was not significant (r = .56, p =
0.28). Most of our stimuli did not have a harmonic structure and
consequently, the determined value of d1 may not reflect the peri-
odicity of the image. As we hypothesized, the value of de appears
more suitable to represent the perceived coarseness.
As for regularity, the results were quite different from those
obtained in the previous experiment. In the previous experiment,
Fig. 7 The relationships
between the scale values and
ACF factors. (a) SV of contrast
and F(0). (b) and (c) SV of
regularity and /1, and de, re-
spectively. (d) SV of coarseness
and de. Arrows in (b) and (c)
indicate two textures mentioned
in the text. Each symbol shows
the SV of each subject
205
we assumed that regular textures have high-amplitude periodical
peaks in the ACF, and that the height of the maximum peak
represents the perceived regularity. The results confirmed these
hypotheses clearly. This experiment, however, failed to yield
similar results. In Fig. 7c, we cannot see a positive correlation
between the SV of regularity and the /1 measure. Rather, there is
a trend indicating a negative correlation in the range of log /1
below –1 (0.1 in real value). This result is highly related to the
difference in the range of /1 value. As can be seen in Table 1, we
used texture images that have quite low values of /1 compared
with Tamura’s experiment. As a result, the maximum peak in the
ACF would not represent the periodicity of the image.
In this case, as hypothesized above, we can use the value of de to
represent the perceived regularity for random textures. We can see
in Fig. 7d that there is a normal correlation between the regularity
and de with two exceptional cases. These two textures indicated in
Fig. 7c-d have values of log /1 above –1. This result can be inter-
preted as follows. When the value of /1 is low, de reflects the
perceived regularity. When /1 exceeds a certain level, however, the
perceived regularity might be affected by the harmonic components
in the texture. It is possible that our subjects perceived the regu-
larity in proportion to the value of de for the textures with less
harmonic structures, and to the value of /1 for the textures with
more harmonic structures.
As in the previous experiment, we applied a multiple regression
analysis to examine whether the plural factors affect the perception
of each property. We found that the factors of F(0) and de were
significant predictors of the SVs of contrast and coarseness (Fig. 8).
The regression coefficients (Table 3) show that the SV of contrast is
represented mainly by F(0). Another predictor d1 in the prelimi-
nary experiment is replaced by de. We could not find any significant
206
Fig. 8 Relationships between predicted SVs by means of multiple
regression analysis and measured SVs of contrast and coarseness.
Each plot represents results for each subject
combination of predictors to fit the SV of regularity. This means
that the other ACF factors are not related to the perceived regu-
larity. At this point, we think that the best predictor of perceived
regularity is de.
Discussion
Considering the results above, we believe that the ACF
analysis is still useful for random textures. The value of
F(0) corresponded well to the perceived contrast, simi-
larly to that in regular textures. The values of F(0) in our
Table 3 Regression coefficients
of F(0) and de and correlation Contrast
coefficients of regression equa-
Subject F(0) de
tions for SVs of contrast and
coarseness
A 0.60** 0.50*
B 0.55* 0.48*
C 0.79** –0.02
D 0.56* 0.47
E 0.57* 0.44
F 0.63** 0.44*
G 0.77** 0.32
H 0.86** 0.08
I 0.74** 0.21
J 0.73** 0.43*
**p <.01, *p <.05
experiment were centered in a narrow range, but were
still found to represent the perceived contrast well. This
result means that the RMS power of the image is con-
sistent with the perceived contrast. The value of de cor-
responded to perceived regularity and coarseness. This
result is consistent with our hypothesis. When a texture
has a harmonic structure, the estimated ACF has peri-
odical peaks corresponding to the periods of the texture.
Both perceived coarseness and perceived regularity are
strongly related to these peaks in the ACF. For the
textures with less harmonic structures, like those used in
this experiment, the initial slope of the ACF would be
important. The value of de shows a degree of correlation
within the image: the more the textures correlate inter-
nally, the more the subjects perceived the coarse and
regular properties. The boundary of these two aspects
for the perception of texture periodicity was the value of
/1  0.1 in this study. The multiple regression analysis
revealed that the perceived contrast and coarseness
could be represented by a linear combination of the
ACF factors F(0) and de with certain weightings. It is
considered that the perceived contrast and coarseness
might have some kind of interaction.
General discussion
We showed that the ACF analysis provides useful
measures for representing the textural properties of
contrast, coarseness, and regularity. This study differs
from the previous studies reviewed in the Introduction
mainly in the following two aspects. First, the examined
textural properties were chosen on the basis of the pre-
vious study of texture similarity judgment (Rao &
Lohse, 1996; Cho et al., 2000). Therefore, there is little
doubt that the observed correspondences between the
textural properties and the computational features are
important for texture modeling. We provided physical
quantities to measure the perceptual strength of texture
properties. Throughout the paper, the validity of the
ACF analysis was examined by comparing the compu-
tational measures with the subjective scores collected for
various kinds of natural textures. Second, the proposed
analysis is driven by a plausible mechanism in the
human visual system. Results imply that extraction of
Coarseness
r Subject F(0) de r
0.85** A 0.26 0.71** 0.80**
0.80* B 0.10 0.55* 0.58
0.78* C 0.23 0.29 0.40
0.79* D 0.49* 0.40 0.68*
0.78* E 0.36 0.63* 0.78*
0.83** F 0.16 0.81** 0.85**
0.89** G 0.09 0.43 0.46
0.88** H –0.49 0.36 0.55
0.80** I 0.05 0.75** 0.76*
0.91** J 0.15 0.70* 0.75*

periodical structure and correlation structure is impor-
tant for perception of natural texture.
As has been previously established, perceived con-
trast could be represented well by the image power.
Multiple regression analysis revealed that perceived
coarseness affected perceived contrast. That is, coarse
textures had higher apparent contrast. The simplest ex-
planation of this observation might be the contrast
sensitivity function of human visual system (Campbell &
Robson, 1968). Human’s spatial contrast sensitivity is
best at mid spatial frequency range around 2–5 c/deg (it
roughly corresponds to the coarse textures in this study).
If it is true for the natural images, the definition of
contrast should include the contrast sensitivity functions
of human observers. However, other studies suggested
that perceived contrast is independent of spatial fre-
quencies as the contrast level is above threshold (Tiip-
pana & Näsänen, 1999). This is called ‘‘contrast
constancy’’. Texture images used in this study are
probably in the range of contrast constancy, because the
subjects could discriminate each image’s internal pattern
reliably. If it holds, perceived contrast should have been
independent of perceived coarseness in this study. This is
an interesting problem arising from our results, but is
beyond the scope of this study.
For representing coarseness and regularity, the ef-
fectiveness of the analysis depended on the structures of
the estimated ACF. When a texture has a harmonic
structure, the estimated ACF has periodical peaks that
correspond to the period of the texture. Perceived
coarseness and regularity were strongly related to the
period and height of these peaks. As for the random
texture, however, the estimated ACF does not have a
periodical structure. In this case, the initial decay rate of
the ACF was important to represent the texture
coarseness and regularity. It seems that there are two
distinct mechanisms in the human visual system to
characterize texture coarseness and regularity. A recent
study also reported that human visual system might use
different cues in discrimination of texture coarseness
depending on the image structure. Sakai & Finkel (1995)
presented evidence that the visual system appears to
track either mean frequency or peak frequency, de-
pending on the strength of harmonic components in the
image. That is, the peak frequency is detected in the
presence of strong peaks, and the mean frequency is
detected in their absence. In the context of our analysis,
the peak frequency and mean frequency in the spectrum
could be comparable to the maximum peak and the
initial decay rate of the ACF.
For the similar purpose as our study, Liu (1997)
measured strength of perceived regularity by use of
spectrum decomposition. As described in the Introduc-
tion, the image was decomposed into the deterministic
(periodic) and indeterministic (random) components in
the frequency domain. Then, the deterministic energy
ratio was calculated for the quantitative measure of
perceived regularity. That is the ratio of the energy
contained in the deterministic component and energy of
207
the full spectrum. Although this method was successful
in calculating perceived regularity significantly, it seems
it did not work well for relatively random textures. It is
probably because the model depends highly on the pe-
riodical structure only. As our results suggested, it needs
to consider both the periodical structure and correlation
structure of the image to represent perceived regularity.
Because the present experiment was based on the com-
parisons within a restricted texture set, it was difficult to
determine a single measure of perceived properties for
both texture sets. Further research to clarify the visual
mechanism is needed for the development of a unified
model of regularity perception for both the periodic and
random textures.
What we have done in this paper is to show how the
ACF parameters can represent texture properties and, to
a certain degree, how human subjects seem to be sensi-
tive to the same parameters captured by the ACF. Thus
we could not make a claim for the ACF mechanism for
human perception. There is, nevertheless, convincing
evidence that such a mechanism exists for visual pro-
cessing (Uttal, 1975; Ben-Av & Sagi, 1995). And what-
ever the mechanism, it is possible that humans do have
access to the correlation structure as extracted by the
ACF. We did not need to make assumptions about the
early filtering stage, but the existence of such a mecha-
nism is obvious. Considering the orientation selective
filters (e.g. Gabor filter), the calculation of the proposed
directional ACF could be possible. Also, the calculation
of the proposed ACF measures are possible as a process
of information characterization and quantification in-
volved after the initial linear filtering stage.
References
Amandasun, M., & King, R. (1989). Textural features corre-
sponding to textural properties. IEEE Transactions Systems,
Man and Cybernetics, 19, 1264–1274.
Badcock, D. R., & Derrington, A. M. (1989). Detecting the dis-
placement of spatial beats: No role for distortion products.
Vision Research, 29, 731–739.
Ben-Av, M. B., & Sagi, D. (1995). Perceptual grouping by simi-
larity and proximity: Experimental results can be predicted by
intensity autocorrelations. Vision Research, 35, 853–866.
Brodatz, P. (1966). Textures. New York: Dover.
Campbell, F. W., & Robson, J. G. (1968). Application of Fourier
analysis to the visibility of gratings. Journal of Physiology
(London), 197, 551–566.
Cho, R. Y., Yang, V., & Hallett, P. E. (2000). Reliability and
dimensionality of judgments of visually textured materials.
Perception & Psychophysics, 62, 735–752.
Cross, G. R., & Jain, A. K. (1983). Markov random field texture
models. IEEE Transactions Pattern Analysis and Machine In-
telligence, 5, 25–39.
Francos, J. M., Meiri, A. Z., & Porat, B. (1993). Unified texture
model based on a 2-D Wold-like decomposition. IEEE Trans-
actions on Signal Processing, 41, 2665–2678.
Hammett, S. T., & Smith, A. T. (1994). Temporal beats in the
human visual system. Vision Research, 34, 2833–2840.
Haralick, R. M. (1979). Statistical and structural approaches to
textures. Proceedings of IEEE, 67, 786–804.
Heeger, D., & Bergen, J. (1995). Pyramid-based texture analysis/
synthesis. In Proceedings of ACM SIGGRAPH.
208
Henning, G. B., Herz, B. G., & Broadbent, D. E. (1975). Some
experiments bearing on the hypothesis that the visual system
analyzes spatial patterns in independent bands of spatial fre-
quency. Vision Research, 15, 887–897.
Julesz, B. (1962). Visual pattern discrimination. IRE Transactions
on Information Theory, 8, 84–92.
Liu, F. (1997). Modeling spatial and temporal textures. PhD thesis,
Massachusetts Institute of Technology.
Liu, F., & Picard, R. W. (1996). Periodicity, directionality, and
randomness: Wold features for image modeling and retrieval.
IEEE Transactions on Pattern Analysis and Machine Intelli-
gence, 18, 722–733.
Malik, J., & Perona, P. (1990). Preattentive texture discrimination
with early vision mechanisms. Journal of Optical Society of
America A, 7, 923–932.
Mao, J., & Jain, A. K. (1992). Texture classification and segmen-
tation using multiresolution simultaneous autoregressive mod-
els. Pattern Recognition, 25, 173–188.
Mayhew, J. E. W., & Frisby, J. P. (1978). Suprathreshold contrast
perception and complex random textures. Vision Resarch, 18,
895–897.
Mosteller, F. (1951). Remarks on the method of paired compari-
sons: III. A test of significance for paired comparisons when
equal standard deviations and equal correlations are assumed.
Psychometrica, 16, 207–218.
View publication stats
Portilla, J., & Simoncelli, E.P. (2000). A parametric texture model
based on joint statistics of complex wavelet coefficients. Inter-
national Journal of Computer Vision, 40, 49–71.
Rao, A. R., & Lohse, G. L. (1996). Towards a texture naming
system: Identifying relevant dimensions of texture. Vision
Research, 36, 1649–1669.
Sakai, K., & Finkel, L. H. (1995). Characterization of the spatial-
frequency spectrum in the perception of shape from texture.
Journal of Optical Society of America, 12, 1208–1224.
Tamura, H., Mori, S., & Yamawaki, T. (1978). Textural features
corresponding to visual perception. IEEE Transactions Systems,
Man and Cybernetics, SMC-8, 460–472.
Thurstone, L. L. (1927). A law of comparative judgment, Psycho-
logical Review, 34, 273–289.
Tiippana, K., & Näsänen, R. (1999). Spatial-frequency bandwidth
of perceived contrast. Vision Research, 39, 3399–3403.
Turner, M. R. (1986). Texture discrimination by gabor functions.
Biological Cybernetics, 55, 71–82.
Uttal, W. R. (1975). An autocorrelation theory of form detection.
Hillsdale, NJ: Erlbaum.
Zhu, S. C., Wu, Y. N., & Mumford, D. (1998). Filters, Random
fields And Maximum Entropy (FRAME) – Towards a Unified
Theory for Texture Modeling. International Journal of Com-
puter Vision, 27, 1–20.