Article

Perception
On Visual Texture 2016, 0(0) 1–25
! The Author(s) 2016
Preference: Can an Ecological Reprints and permissions:
[Link]/[Link]

Model Explain Why People DOI: 10.1177/0301006616629026

[Link]

Like Some Textures More

Than Others?
Kyle D. Stephens and Donald D. Hoffman
University of California, Irvine, CA, USA

Abstract
What visual textures do people like and why? Here, we test whether the ecological valence theory
proposed for color preferences can also predict people’s preferences for visual texture. According
to the theory, people should like visual textures associated with positive objects or entities and
dislike visual textures associated with negative objects or entities. We compare the results for the
ecological model with a more traditional texture-preference model based on computational
features and find that the ecological model performs reasonably well considering its lower
complexity, explaining 63% of the variance in the human preference data.

Keywords
Visual texture, aesthetics, preference, ecological valence theory

Introduction
For members of the species H. sapiens, textures are ubiquitous. Every day we are surrounded
by objects and surfaces that give us unique feelings when we touch them—fur feels soft, tree
bark feels rough, and silk feels smooth. These different feelings arise from variations on the
surfaces of objects, which also lead to variations in the pattern of light that reaches our
eyes—fur looks soft, tree bark looks rough, and silk looks smooth. We call this visual aspect
of surface variation visual texture.1 Artists and designers have long used visual texture as a
tool to evoke emotions and set moods (see, e.g., Brodatz, 1966; Gatto, Porter, & Selleck,
2010). In the current article, we investigate the aesthetics of visual texture more formally by
empirically exploring the question: What visual textures do people like and why? We examine
whether the ecological valence theory (EVT) proposed by Palmer and Schloss (2010) for color
preferences—and later extended to odor preferences (Schloss, Goldberger, Palmer, &

Corresponding author:
Kyle D. Stephens, University of California, 3151 Social Sciences Plaza, Irvine, CA 92697-5100, USA.
Email: kdstephe@[Link]
2 Perception 0(0)

Levitan, 2015)—can also be used to explain preferences for visual texture. But first we review
previous work on visual texture and the aesthetics of images.
There are many models that predict perceptual properties of visual textures. For example,
there are models that predict how rough (Ho, Landy, & Maloney, 2006; Tamura, Mori, &
Yamawaki, 1978), glossy (Anderson & Kim, 2009; Kim & Anderson, 2010; Montoyishi,
Nishida, Sharan, & Adelson, 2007), or complex (Amadasun & King, 1989; Tamura et al.,
1978) they look, as well as how they are segmented (Ben-Shahar, 2006; Julesz, 1962, 1981;
Landy & Bergen, 1991; Malik & Perona, 1990; Portilla & Simoncelli, 2000; Rosenholtz, 2000;
Tyler, 2004; Victor, 1988) or classified (e.g., Dong, Tao, Li, Ma, & Pu, 2015; Guo, Zhang, &
Zhang, 2010; Haralik, Shanmugam, & Dinstein, 1973; Randen & Husy, 1999; Varma &
Zisserman, 2005) by human observers (for reviews, see Bergen, 1991; Landy & Graham,
2004; Rosenholtz, 2015).
There are also many models that predict aesthetics of images, including how much
observers like the images (e.g., Chen, Sobue, & Huang, 2008; Kawamoto & Soen, 1993; S.
Kim, Kim, Jeong, & Kim, 2006; Lee & Park, 2011; Um, Eum, & Lee, 2002; Zhang et al.,
2011; for reviews, see Joshi et al., 2011; Palmer, Schloss, & Sammartino, 2013). But there are
not many models that focus on the aesthetics of visual textures specifically, and current
models do not necessarily extend well to the general space of visual textures (Thumfart
et al., 2011). To our knowledge, there is only one extant model that specifically predicts
preference for visual textures; it’s based on computational features and ratings of other
aesthetic properties.

Current Computational Model of Visual Texture Aesthetics

Thumfart et al. (2011) and Liu, Lughofer, and Zeng (2015) have both recently published a
parametric model that predicts human ratings on several aesthetic properties2 for visual
textures (Figure 1). Given a visual texture, this model makes predictions using a
combination of low-level computational features measured from that texture and human
ratings on other aesthetic properties for that texture. For instance, Thumfart et al. found
that how elegant a texture looks was best predicted by a combination of eight computational
features measured from that texture (top three: the response energy of a particular oriented
Gabor filter applied to the image, the contribution of blue to the image color, and the mean
variance of image intensities) and human ratings on four other aesthetic properties for that
texture (ratings of how glossy, silky, woodlike, and grainy it looks). In general, the
computational features are derived from well-known image-processing features—like the
Fourier power spectrum of an image or the gray-level co-occurrence matrix for an
image—and the aesthetic properties include primary/physical descriptions (e.g., Cold–
Warm, Rough–Smooth), more abstract descriptions (e.g., Simple–Complex, Inelegant–
Elegant), and preference (Dislike–Like).
This model is hierarchical in that some aesthetic properties are predicted using only
computational features, whereas others are predicted using a combination of
computational features and other, lower-level, aesthetic properties (see Figure 1). Both
Thumfart et al. and Liu et al. consider preference to be the highest level aesthetic
property, and thus their models predict preference using all other computational features
and aesthetic properties.3 For instance, Thumfart et al. predicted preference using linear
regression with 214 potential regressors: 188 computational features and ratings for all 26
other aesthetic properties (see Figure 1).
Stephens and Hoffman 3

Figure 1. Structure of the hierarchical feed-forward model of texture aesthetics used by Thumfart et al.
(2011) and Liu et al. (2015; adapted from Thumfart et al.’s representation of the model). The model predicts
human ratings of aesthetic properties using measurements of computational features and ratings of other
aesthetic properties. In the current article, we are interested in preference, or Dislike–Like.

This model has been able to predict human preference for visual texture with high
accuracy (in Liu et al.’s instantiation of the model, they claim to account for as much as
99% of the variance in human preference ratings). However, the model employs a see-what-
sticks approach that emphasizes model prediction over model understanding. The numerous
features and aesthetic properties constitute a veritable grab bag from the literature on
computer vision, image processing, and aesthetics. But, to the credit of Thumfart et al.
and Liu et al., their models are more interpretable than black box models such as neural
networks, and Thumfart et al. increased interpretability by using an optimization measure
that punishes complexity (defined by number of regression terms).
Still, these models were not constrained by any overarching theory dictating which
features should be included and which should not, and it’s unclear why certain features
are so important for predicting preference while others are not. For instance, Thumfart
et al. found that how premium, sophisticated, and woodlike a texture looks were all
important for predicting how much someone likes that texture, whereas how elegant and
natural it looks were not. Likewise, they found that the skewness of the energy contained in
12 circular ring segments of the Fourier power spectrum of a texture was important for
predicting how much someone likes that texture, whereas the kurtosis or standard
deviation of the energy contained in those same 12 ring segments was not (see Table 1).
While these results are interesting and useful, they are not theoretically motivated and would
not have been predicted by any current theory.
In other words, this model is good at predicting what visual textures people like but does
not give the best sense of why.
4 Perception 0(0)

Table 1. Tested Models and Hypotheses.

Preference prediction
Model/hypothesis equation Variables explained
P
Full EVT model Pt ¼ n1t wtd vd Pt: predicted preference for texture t
d2Dt nt: number of descriptions ascribed to texture t
Dt: set of all descriptions ascribed to texture t
wtd: average match weight between texture t and
description d
vd: average valence rating for description d

Single-associate Pt ¼ vd0 ðtÞ vd0 ðtÞ : average valence rating for description d0
hypothesis (restricted d0 (t): description most frequently ascribed to
EVT model) texture t

Namesake hypothesis Pt ¼ vd ðtÞ vd ðtÞ : average valence rating for description d*
(restricted EVT model) d*(t): description of what t is actually a texture of

Thumfart et al. model Pt / premium þ premium, sophisticated, rough, woodlike:

sophisticated þ aesthetic properties gatherd for texture t via
rough þ cSkew þ human ratings.
woodlike þ strength cSkew, strength: computational features derived
from the Fourier power spectrum and the
neighborhood gray tone difference matrix for
texture t, respectively.

Note. EVT ¼ ecological valence theory.

Ecological Valence Theory

In contrast to the see-what-sticks computational model, Palmer and Schloss’ (2010)
ecological valence theory is based on a simple premise: People’s preferences for basic
stimuli (e.g., colors, odors) are derived from how they feel about the objects or entities
associated with those stimuli (note that the theory was initially proposed only for colors
and only later extended to other types of stimuli/modalities). So, for example, a student might
have a particular affinity for the color blue because they like blueberries or attend the
University of California, Irvine (whose official color is blue), whereas another student
might have an affinity for the color red because they like cherries or attend the University
of Southern California (whose official color is red). Indeed, there is evidence that people like
the colors of their own social group more than members of a rival social group (Schloss &
Palmer, 2014; Schloss, Poggesi, & Palmer, 2011).
But, Palmer and Schloss take this basic premise a step further and claim that people’s
preferences are dictated by a summary statistic that accounts for the valence across all objects
associated with a particular stimulus. That is, how much a person might like a given stimulus
can be explained by how positive or negative people feel, on average, about all objects and
entities associated with that stimulus. Thus, for example, a student might like the color blue
not just because it is associated with UC Irvine, but because it is generally associated with a
multitude of positive objects (e.g., clear sky, clean water, blue jeans) and not many negative
objects. The motivation for this reasoning is evolutionary: Assuming that people have
Stephens and Hoffman 5

positive feelings about objects that lead to beneficial outcomes, then stimuli that are typically
associated with those objects serve as indicators to the beneficial outcomes. Thus, such
preferences should steer organisms to approach objects that lead to beneficial outcomes and
avoid those that lead to harmful outcomes, an evolutionarily advantageous strategy (Palmer &
Schloss, 2010; Schloss, Goldberger, et al., 2015). This is the ecological valence theory.
Palmer, Schloss, and colleagues have already found support for this theory for colors and
odors: Average color preferences are well predicted by how people feel about all objects
associated with the colors (Palmer & Schloss, 2010; Taylor & Franklin, 2012), and average
odor preferences are well predicted by how people feel about all objects associated with the
odors (Schloss, Goldberger, et al., 2015). Furthermore, the theory predicts that preferences
across cultures or social groups should vary only to the degree that these groups associate
different objects with the same stimuli or have different feelings about the same objects, and
there is empirical support for this with colors (Schloss, Hawthorne-Madell, & Palmer, 2015;
Yokosawa, Yano, Schloss, Prado-Leon, & Palmer, 2010).
Thus far, Palmer, Schloss, and colleagues have only tested the EVT with colors and odors,
but the theory should be generally applicable to different kinds of stimuli in any modality
(see, e.g., Schloss, Goldberger, et al., 2015). The goal of the present research was to test how
well the EVT accounts for preferences for visual texture: Can people’s average preferences for
visual textures be predicted by how positive or negative they feel about all objects associated
with the textures?

Present Study
Testing the ecological valence hypothesis. To test whether preference for a given visual texture can
be predicted by the average valence for all objects or entities associated with that texture, we
adapted the procedure of Palmer and Schloss (2010) to visual textures. This procedure
required data from four different groups of participants.
The first group gave baseline preference ratings which we used to evaluate model
performance. These participants were simply asked to rate how much they liked each
visual texture on a line-mark ratings scale coded from 100 (dislike) to þ100 (like).
The next three groups of participants were needed to calculate the predicted preference4
for each texture (see Tables 1 and 2): One group typed out descriptions of objects they
associated with each visual texture (the object-description group); another group rated
their emotional response (positive/negative) to a condensed set of these object descriptions
(the description-valence group); and the final group rated how well each visual texture
matched all of the descriptions ascribed to that texture (the texture-description match
group). The prediction for any visual texture was then given by the sum of the average
valence ratings for each object associated with the texture, weighted by how well those
objects match the texture:
1X
Pt ¼ wtd vd ð1Þ
nt d2D
t

where, Pt is the predicted preference for texture t (100 to 100), nt is the number of
descriptions ascribed to texture t, Dt is the set of all descriptions ascribed to texture t, wtd
is the average match weighting between texture t and description d (0 to 1), and vd is the
average valence rating for description d (100 to 100). Using this formulation, Palmer and
Schloss (2010) accounted for 80% of the variance in average color preferences, and Schloss,
Goldberger, et al. (2015) account for 76% of the variance in average odor preferences.
6 Perception 0(0)

Table 2. Participants and Their Experimental Tasks.

Number of
Participant participants Data from task (see
group Total (used) Task Description of task variables in Table 1)

1 36 (36) Texture-preference Participants rated how much Baseline ratings for

rating they like each visual model checking
texture on a scale from
Dislike (100) to Like
(þ100).
2 32 (31) Object description Participants typed out nt and Dt for all
descriptions of objects textures t
that they thought might
describe or be associated
with each visual texture.
3 37 (37) Description- Participants rated their vd for all descriptions d
valence rating emotional response to
each object description
(consolidated from Group
2) on a scale from
Negative (100) to
Positive (þ100).
4 27 (26) Texture-description Participants rated how well wtd for all textures t
match rating the descriptions and descriptions d
(consolidated from Group
2) matched the visual
textures they were
ascribed to on a scale from
Very Poorly (0) to Very
Well (1).
5 24 (24) Aesthetic-property Participants rated visual Ratings for premium,
rating textures on the scales sophisticated, and
NotPremium–Premium, rough
NotSophisticated–
Sophisticated, and
Smooth–Rough (all 100
to þ100).
6 22 (21) Aesthetic-property Participants rated visual Ratings for woodlike
rating textures on a scale ranging
from NotWood (100) to
Wood (þ100).

Testing other hypotheses. One might (reasonably) question whether applying this formulation to
visual textures is sensible. After all, isn’t it generally obvious what object a (natural) visual texture
corresponds to? In this case, preference for a visual texture would reduce to how people feel
about the single associated object, an uninteresting result. Schloss, Goldberger, et al. (2015) faced
a similar problem when applying the EVT to odors because they used odor pens that were
designed to smell like a particular object (e.g., honey, apple, leather). But they found that the
associated objects were more ambiguous than one would think and that odor preference was
better explained by the valence for all objects associated with the odor pens, rather than just the
object the pen was designed to smell like. We hypothesized a similar outcome for visual textures.
Stephens and Hoffman 7

To test this outcome, we examined two additional hypotheses (following Schloss,

Goldberger, et al., 2015): the single-associate hypothesis and the namesake hypothesis
(see Table 1). The single-associate hypothesis states that preference for a visual texture
is best predicted by the valence of the single object that the texture is most associated
with. The namesake hypothesis states that preference for a visual texture is best predicted
by the valence of the namesake object that produced the visual texture. So, for example,
one of the visual textures used was a close-up of a strawberry (see Figure 2). The EVT
predicts that preference for this texture is best explained by the weighted average valence
ratings for all descriptions ascribed to this texture (e.g., ‘‘beans,’’ ‘‘flowers,’’ ‘‘eggs,’’
‘‘peacock tail,’’ ‘‘seeds,’’ ‘‘strawberry,’’ ‘‘insect eyes,’’ etc.); the single-associate
hypothesis predicts that preference for this texture is best explained by the average
valence of the description ‘‘eggs’’ because this was the description most often associated
with the texture; and the namesake hypothesis predicts that preference for this texture is
best explained by the average valence of the description ‘‘strawberry’’ because this is the
namesake object that the texture was produced from.

Comparing with the computational models. After testing the single-associate and namesake
hypotheses, we also compared the results of the ecological valence model with the
computational model of Thumfart et al. (2011). First, we compared the error of the
ecological model with the error of Thumfart et al.’s model using an error term that
punishes the models for complexity (see Equation 2 and Appendix B). Then, we tested
Thumfart et al.’s model more directly.
Thumfart et al. found that preference for a given visual texture was best predicted by a
linear equation with six factors: two computational features and four aesthetic properties (see
Table 1). Thus, to test Thumfart et al.’s prediction equation, we calculated the computational
features for each texture, and we recruited participants to rate the textures on a line-mark
scale for each aesthetic property.

Review. Table 1 gives an overview of the different models and hypotheses tested in this article.

Figure 2. The 10 visual textures from Shutterstock used for all tasks. The rightmost pictures are Gaussian-
blurred versions of the preceding pictures. From top left to bottom right: lettuce, strawberry, tropical ocean,
chocolate, Gaussian-blurred chocolate, mold, rotted strawberry, skin infection, mud, and Gaussian-
blurred mud.
8 Perception 0(0)

Materials and Methods

Participants
There were six separate groups of participants, all of whom were undergraduate students at
the University of California, Irvine and received course credit for participation. No
participant ever participated in more than one experimental task (see Table 2 for an
overview of the six groups of participants and their different tasks).
Thirty-six participants (19 men) took part in the texture-preference rating task, rating how
much they liked each visual texture. The mean age was 20.8 years (range 18–28 years).
Thirty-two participants (15 men) took part in the object-description task, describing
objects associated with each visual texture. Once all of the descriptions were consolidated
(see Procedure section), one of the authors (K. Stephens) rated how well he thought the
descriptions matched the visual textures they were ascribed to on a 0 to 100 line-mark scale.
We ended up excluding one female participant because 38 of her descriptions were given a
0 match rating by this author (this was 4.5 standard deviations above the mean number of
0 match ratings for the rest of the group). After excluding this participant, the mean age was
21.6 years (range 18–37 years).
Thirty-seven participants (18 men) took part in the description-valence rating task, rating
how positive or negative they felt about each object description. The mean age was 20.7 years
(range 18–32 years).
Twenty-seven participants (14 men) took part in the texture-description match rating task,
rating how well each description matched the texture it was ascribed to. We excluded one
male participant because he did not finish the experiment (we lost connection to our
MATLAB license three quarters of the way through). After excluding this participant, the
mean age was 22.5 years (range 18–40 years).
Twenty-four participants (11 men) took part in the first aesthetic-property rating task,
rating how premium, sophisticated, and rough the textures looked. The mean age was 21.1
years (range 18–36 years).
Finally, 22 participants (6 men) took part in the second aesthetic-property rating task,
rating how woodlike the textures looked. We excluded one female participant because
her mean reaction time was more than two standard deviations faster than the mean for
the group. After excluding this participant, the mean age was 20.7 years (range 18–34
years).

Stimuli
Texture-preference rating, object description, and aesthetic-property rating. For these tasks, the stimuli
consisted of 62 different visual textures. Fifty-two of these textures were from Brodatz’s
(1966) album, and 10 were from [Link]. We used the same Brodatz textures as
those used by Rao and Lohse (1993, 1996), excluding D43, D75, D102, and D108 (see Tables
A1 and A2 in Appendix A for a complete list). Of the Shutterstock images, eight were chosen
specifically to have either high valence (chocolate, fresh lettuce, fresh strawberry, sea/ocean)
or low valence (mud, mold, rotten strawberry, infected skin); the remaining two images were
Gaussian-blurred versions of the chocolate and the mud images, respectively. All images were
grayscale and viewed at 256-by-256 pixels (see Figure 2 for the Shutterstock images; see
Tables A1 and A2 in Appendix A for a complete list of all visual textures; see Brodatz,
1966 or Rao & Lohse, 1993 for pictures of the Brodatz textures).
Stephens and Hoffman 9

Texture-description match rating. Because of the large number of descriptions for each texture
(even after consolidation; see Procedure section), it would have taken too long to show
participants all texture-description pairs for all 62 textures in the texture-description match
rating task. Thus, we included only 47 of the original 62 textures in this task. We included all
10 of the [Link] images, but only 37 of the original 52 Brodatz textures. We chose
which 15 Brodatz textures to exclude (D11, D25, D29, D31, D39, D40, D42, D55, D63, D80,
D82, D83, D89, D94, D101) based on their similarity to other textures already included (see
Table A1 in Appendix A for a complete list of all visual textures used in this experiment).

Description-valence rating. For the description-valence rating task, the stimuli consisted of black
text presented in 28-point Arial font.

Displays
Monitor for visual textures. For any task displaying visual textures (i.e., the texture-preference
rating, object description, texture-object match rating, and aesthetic-property rating tasks),
stimuli were presented using a 27-inch Apple iMac display set at a resolution of 2,560  1,440
and a refresh rate of 60 Hz. All stimuli were presented against a neutral gray background with
luminance 28.8 cd/m2 (as measured using Photo Research PR-670 spectroradiometer). The
monitor was calibrated using an X-Rite ColorMunki Photo spectrophotometer. Participants
were all run in the same room, one at a time.

Monitor for descriptions. For the task displaying only text (i.e., the description-valence rating
task), stimuli were presented using Dell computers attached to 17-inch Dell LCD monitors
set at a resolution of 1,280  1,024 and a refresh rate of 60 Hz. The text stimuli were displayed
against a white background. Participants were run in groups of one to five in a room
containing eight computers.

Line-mark rating scale. Five of the tasks—the texture-preference rating, description-valence

rating, texture-description matching rating, and aesthetic-property rating tasks—gathered
ratings using a line-mark rating scale, located at the bottom of the display. This scale
consisted of a dark gray rectangular box demarcated into 10 equally spaced regions (5
below the center point and 5 above), with text above the scale indicating what rating was
being performed (e.g., ‘‘How much do you like this texture?’’), text at the end points of the
scale to indicate the extreme ratings (e.g., ‘‘not at all’’ vs. ‘‘very much’’), and a black slider bar
that participants could slide along the scale with the mouse to indicate their rating.
Participants could either click and drag the slider bar or move the bar by pointing and
clicking (i.e., without dragging). When displayed on the larger monitor (i.e., for the
texture-preference rating, texture-object match rating, and aesthetic-property rating tasks),
the top edge of the question text was approximately 247 pixels below the bottom edge of the
texture, the rating scale was 768 pixels wide by 72 pixels high, and the scale’s top edge was
approximately 41 pixels below the bottom edge of the question text. When displayed on the
smaller monitor (i.e., for the description-preference rating task), the rating scale was made
smaller—384 pixels wide by 52 pixels high—in proportion to the smaller resolution of the
screen. The black slider bar was always 6 pixels wide by the height of the scale (either 72 or 52
pixels), and all text was black, in 24-point Arial font. The scale and the question text were
always horizontally centered. The code for all tasks using a line-mark rating scale was written
in MATLAB, using the Psychophysics Toolbox extensions (Brainard, 1997; Kleiner,
Brainard, & Pelli, 2007).
10 Perception 0(0)

Description edit boxes. The object-description task gathered descriptions using edit boxes
located below the textures. There were five edit boxes in total, where participants could
click and enter text. The edit boxes were 200 pixels wide by 50 pixels high and were
located approximately 137 pixels below the bottom edge of the texture. They were
horizontally centered and evenly spaced approximately 123 pixels apart from one another.
Below the edit boxes was a pushbutton labeled ‘‘submit’’ where participants could click to
input their responses and move on to the next trial. This pushbutton was 150 pixels wide by
50 pixels high, horizontally centered, and the top edge was approximately 96 pixels below the
bottom edge of the edit boxes. The code for this task was written using MATLAB’s uicontrol
functions.

Procedure
The present study included five between-subject tasks: (a) texture-preference rating, (b) object
description, (c) description-valence rating, (d) texture-description match rating, and (e)
aesthetic-property rating (see Table 2). For all tasks, there was a blank 500 ms ISI in
between trials, and participants sat approximately 60 cm away from the monitor. At this
viewing distance, the visual textures spanned approximately 5.7 degrees of visual angle.

Texture-preference rating task. On each trial, participants viewed one of the 62 visual textures,
vertically and horizontally centered on a neutral gray background with the text ‘‘How much
do you like this texture?’’ displayed below the texture. Participants indicated their preference
using a line-mark scale labeled ‘‘not at all’’ on the left end (coded as 100) and ‘‘very much’’
on the right end (coded as þ100). Participants used the mouse to indicate their preference.
They pressed the spacebar to input their rating and move on to the next texture. The textures
were displayed in random order, and all participants gave ratings for all textures.

Object-description task. Participants saw each texture individually against a neutral gray
background and were instructed to type out as many objects as they could think of (up to
five) that might describe or be associated with the given visual texture. They were instructed
not to name adjectives (e.g., ‘‘heavy,’’ ‘‘light,’’ ‘‘furry’’), abstract entities (e.g., ‘‘dreams,’’
‘‘hopes,’’ ‘‘fears’’), or objects that would not be known to other people (e.g., ‘‘my aunt’s
blanket’’). They were also told that the experimenters were interested in all associated objects,
whether pleasant or unpleasant. They typed their responses into one of five edit boxes located
below the texture on the screen and clicked a pushbutton labeled ‘‘submit’’ to move on to the
next texture once they were satisfied with their descriptions. They were not given a time limit.
The textures were displayed in random order, and all participants gave descriptions for all 62
textures.
After combining exact repeats, we were left with 1,264 unique descriptions (from an initial
total of 4,858). We compiled these descriptions into a single list of 206 items using a
procedure similar to that of Palmer and Schloss (2010). First, we discarded items from the
list if they (a) were abstract concepts instead of objects (e.g., ‘‘anger,’’ ‘‘dream,’’ ‘‘joy’’); (b)
were adjectives instead of objects (e.g., ‘‘big,’’ ‘‘bumpy,’’ ‘‘blurry,’’ ‘‘curvy,’’ ‘‘dark,’’
‘‘heavy’’); or (c) were ambiguous or could describe many different things (e.g.,
‘‘arrangement,’’ ‘‘game,’’ ‘‘layers,’’ ‘‘material,’’ ‘‘mixture,’’ ‘‘surface,’’ ‘‘painting’’). We
excluded 366 descriptions this way, leaving 898 items on the list. We then categorized the
descriptions to reduce the number that needed to be rated in the description-preference rating
and the texture-object match rating tasks. We combined descriptions into a single object
category if they seemed to refer to the same object (e.g., ‘‘blanket,’’ ‘‘quilt,’’ and ‘‘piece of a
Stephens and Hoffman 11

blanket’’; ‘‘beehive,’’ ‘‘bee’s nest,’’ and ‘‘bee cells’’; ‘‘carpet,’’ ‘‘carpret [sic.],’’ and ‘‘rug’’),
and, in certain cases, we combined descriptions into superordinate categories if there were
several exemplars referring to the same type of object. In the latter case, exemplars were
included in the category description (e.g., ‘‘bedding [covers, sheets, pillows]’’; ‘‘bricks [brick
wall, brick building, brick sidewalk]’’; ‘‘reptile skin [snake, crocodile, lizard]’’). Descriptions
were further excluded if they were given by only one person for only one image and they
didn’t fit into any other object categories. After this process, we ended up with the final list of
206 consolidated object descriptions.
Even after this consolidation, each texture still had, on average, 20.4 descriptions
attributed to it. With this many texture-description pairings, the texture-description
matching task would have been too long (even after pairing the textures down from 62 to
47), so two independent raters ran a preliminary version of the texture-description match
rating task, and we excluded any texture-descriptions pairings that were given by only one
participant and assigned a 0 match rating by both independent raters. After this procedure,
the textures had an average of 18.4 descriptions each.

Description-valence rating task. Following Palmer and Schloss (2010), participants were first
given a list of eight sample items (sunset, bananas, diarrhea, sidewalk, boogers/snot, wine
(red), chalkboard, chocolate) to give them an idea of the range of descriptions they would see.
They were then presented with each of the 206 consolidated descriptions in black text,
horizontally and vertically centered against a white background. The procedure was the
same as for the texture-preference rating task, but the instruction text was different. Below
each description was the text, ‘‘What’s your emotional reaction to the object(s)/thing
described above?’’ and participants indicated their rating using a line-mark rating scale
labeled ‘‘negative’’ on the left end (coded as 100) and ‘‘positive’’ on the right end (coded
as þ100). The descriptions were displayed in random order, and all participants gave ratings
for all descriptions.

Texture-description match rating task. Because of the large number of texture-description

pairings, we used only 47 of the original 62 visual textures in this task to keep it from
being too long. We chose which textures to exclude based on their similarity to other
textures already included (see Stimuli section). On each trial, participants viewed one of
the 47 textures along with one of the descriptions ascribed to that texture against a neutral
gray background. The description was displayed in black, 28-point Arial font, horizontally
centered, approximately 110 pixels above the top edge of the texture. Below each texture-
description pairing was a line-mark scale with the text ‘‘How well does this description match
the image?’’ The scale was labeled ‘‘very poorly’’ on the left end (coded as 0) and ‘‘very well’’
on the right end (coded as 1). Participants saw all possible parings of texture images and
descriptions ascribed to those textures. We split the participants for this experiment into two
groups: One group saw the same description for all textures given that description before
moving on to the next description, and the other group saw the same texture for all
descriptions given to that texture before moving on to the next texture. The order of the
descriptions and the textures was randomized.

Aesthetic-property rating tasks. Participants viewed each of the 62 visual textures, one at a time,
against a neutral gray background with the text ‘‘This texture looks:’’ displayed below the
texture. They indicated their rating for the requested aesthetic property using a line-mark
scale coded from 100 to þ100.
12 Perception 0(0)

For the first aesthetic-property rating task, participants rated the textures on three different
antonym pairs: NotPremium–Premium, NotSophisticated–Sophisticated, and Rough–Smooth.
The experiment was blocked so that participants rated all textures on one aesthetic property
before moving on to the next, and they were given a short break between blocks. The order of
the aesthetic properties was randomized as was the order of the textures within a block.
The second aesthetic-property rating experiment was identical to the first, except that
participants rated the textures on only one antonym pair: NotWood–Wood.

Review. Table 2 gives an overview the six groups of participants and their experimental tasks.

Results and Discussion

We first describe participants’ texture-preference ratings for the 47 visual textures used in all
tasks. We then test the ecological valence hypothesis for these textures, namely that texture
preferences can be explained by the combined valence of all objects associated with the
textures. Next, we test this account against the single-associate and namesake hypotheses.
Finally, we fit the model of Thumfart et al. (2011) and compare the results of this model with
those of the ecological valence model.

Texture Preferences
We used Cronbach’s coefficient alpha (Cronbach, 1951; see also Ritter, 2010) to assess
interrater reliability for texture-preference ratings. Reliability was high (.83), so we
averaged across participants to get mean preference ratings. Figure 3 shows the mean
preference ratings for the 47 visual textures used in all experiments (see Table A2 in the
Appendix A for preference ratings for the other 15 visual textures used in this experiment).
The most preferred textures included clouds, lettuce, lace, and tropical sea. The least
preferred textures were the two blurred images, mud, and rotted strawberry.

Testing the Ecological Account of Texture Preferences

We used Cronbach’s coefficient alpha to assess interrater reliability separately for the groups
that rated description-valence and description-texture matching. Reliability was high for
these groups (.96, .99), so we averaged across participants to obtain mean valence-ratings
for each description (vd) and mean match ratings6 for each texture-description pairing (wtd).
Using these ratings along with the consolidated descriptions from the object-description task,
we were able to predict preference ratings for the each texture using Equation 1.
Figure 4 shows the strong positive correlation (r ¼ .79, p < .001) between the measured
texture preferences, averaged across participants, and the preferences as predicted by
Equation 1. This indicates that people like textures that remind them of positive objects
and dislike textures that remind them of negative objects, and it supports the ecological
valence hypothesis. It is also worth noting that the prediction equation accounted for less
variance if we either dropped the normalizing term 1/nt (r ¼ .66), or set all of the nonzero
weights (wtd’s) to 1 (r ¼ .72).
The main problem with the predicted ratings is that their range is compressed compared
with the measured ratings and that most of the predicted ratings are positive while many of
the measured ratings are negative. Palmer, Schloss, and colleagues reported similar issues
when predicting color preferences (Palmer & Schloss, 2010) and odor preferences (Schloss,
Goldberger, et al., 2015). To explain why the model underpredicts negative ratings, they
Stephens and Hoffman 13

Figure 3. Average texture-preference ratings (y axis) for each of 47 visual textures (x axis). For Brodatz
textures, Brodatz’s (1966) identifier is given in parentheses, and the labels are his descriptions. For the other
textures, the labels come from the Shutterstock search terms that produced the images. Error bars represent
the standard errors of the means.

hypothesized that participants in the object-description group either underreport negative

objects because they are too shy to report gross or disgusting things, or they are just biased
toward generating/thinking about positive objects. To explain why the ratings for the
prediction are compressed relative to the measured ratings, we posit that people likely
have weaker feelings about text describing objects than they have about actual images.
Despite these failings, the model does reasonably well: The predictions account for 62.5%
of the variance in measured preference with a single predictor.

Evaluating the Single-Associate Hypothesis

We tested the single-associate hypothesis by first identifying the object named most
frequently for each visual texture (see Table A1 in Appendix A). We then calculated the
correlation between average preference ratings for the 47 visual textures and average valence
ratings for the most frequently named objects. Although there was a significant positive
correlation (r ¼ .47, p < .001), it was weaker than the relation between measured
preferences and preferences predicted by Equation 1 (correlations compared using the
Fisher r-to-z transformation: z ¼ 2.63, p ¼ .01). Therefore, texture preferences were better
predicted by a summary statistic of valence for all objects associated with a texture than
by valence for the single object most frequently associated with a texture.
14 Perception 0(0)

Figure 4. For a set of 47 visual textures, shows the correlation (r ¼ .79, p < .001) between measured
preferences, averaged across participants, and the preferences predicted by Equation 1. Numbers represent
ratings on a line-mark rating scale, where 100 indicates that the texture is liked ‘‘not at all,’’ and þ100
indicates that it is liked ‘‘very much.’’

It is also worth noting that, although the valence of the most frequent description did not
predict texture preference better than the valence of all associated descriptions, texture
preference was negatively correlated with the number of descriptions (r ¼ .67, p < .001) so
that the textures with fewer descriptions were better liked. Taylor and Franklin (2012) found
this to be the case with colors as well, but Schloss, Hawthorne-Madell, & Palmer (2015) did
not, nor did Schloss, Goldberger, et al. (2015) find this with odors. On average, each texture
had 18.4 descriptions associated with it (range ¼ 6–34).

Evaluating the Namesake Hypothesis

We tested the namesake hypothesis by first identifying the ‘‘namesake object’’ for each visual
texture, that is, the object from which the texture image was created. We then calculated the
correlation between average preference ratings for the visual textures and average valence
ratings for the namesake objects.
To identify the namesake objects, we chose the consolidated description that was closest
either to Brodatz’ (1966) description (for the Brodatz textures) or to the Shutterstock search
term that produced the image (for the Shutterstock textures). The matches were not always
exact but were reasonably close. For example, Brodatz described texture D37 as ‘‘water,’’ but
we used ‘‘ocean/sea’’ as the namesake object. Similarly, one of the Shutterstock textures was
produced from the search term ‘‘rotten strawberry,’’ but we used ‘‘rotten/spoiled food’’ as the
namesake object (see Table A1 in Appendix A for more details). We excluded texture D32
because nothing close to Brodatz’ description for the texture (pressed cork/corkboard) was
ever given as an object description, and thus D32’s namesake was never given a valence
Stephens and Hoffman 15

rating. We also excluded the two Gaussian-blurred visual textures because the namesake
objects for these textures would be too difficult to visually determine. Thus, overall, the
correlation was based on ratings for 44 of the visual textures.
Although there was a significant positive correlation (r ¼ .60, p < .001), it was weaker than
the relation between measured preferences and preferences predicted by Equation 1 (with the
44 textures: r ¼ .82, p < .001; correlations compared using the Fisher r-to-z transformation:
z ¼ 2.10, p ¼ .04). Thus, texture preferences were better predicted by the average valence over
all objects associated with that texture than by the valence of the namesake object that
produced the texture.7

Comparing with the Computational Model of Thumfart et al.

While the ecological model accounted for 62.5% of the variance in texture-preference ratings,
the accuracy of this prediction was not nearly as good as that provided by the computational
model of Thumfart et al. (2011) and Liu et al. (2015), the latter of which claimed to account
for as much as 99% of the variance in human preference ratings. However, the ecological
model used only a single predictor, whereas the models of Thumfart et al. and Liu et al. used
six and five predictors, respectively. Because Thumfart et al. optimized their model using an
error measure that punished the model for complexity (i.e., number of predictors), we were
able to compare the results of the ecological model with the results of Thumfart et al.’s
model, taking into account the lower complexity of the ecological model. The punished
error measure was a convex combination of the following form (see Appendix B for more
details):
Errorpunished ¼   Accuracy þ ð1  Þ  Complexity ð2Þ

with  in the range [0,1]. Thus, the parameter  controls the ratio of accuracy versus
complexity, with lower complexity emphasized for decreasing values of . Using this
formulation, we found that the punished error for the ecological model matched that of
Thumfart et al.’s model when  ¼ .35 (see Appendix B). Thus, although the ecological
model is less accurate than Thumfart et al.’s model, it has comparable performance when
the modeler values low complexity over high accuracy. We were unable to make such a
comparison with Liu et al.’s model because they did not work with a punished error term.
We also tested Thumfart et al.’s model more directly by measuring the six factors in their
final texture-preference prediction equation (see Table 1): (a) how premium a texture looks,
(b) how sophisticated a texture looks, (c) how rough5 a texture looks, (d) how woodlike a
texture looks, (e) the skewness of the distribution of Fourier energy in concentric rings in the
Fourier power spectrum of a texture (cSkew), and (f) the strength, a computational feature
derived from the neighborhood graytone difference matrix of a texture. We gathered ratings
for premium, sophisticated, rough, and woodlike in the aesthetic-property ratings tasks, and
we measured the computational features cSkew and strength following Thumfart et al. (2011)
and Amadasun and King (1989), respectively. We used Cronbach’s coefficient alpha to assess
interrater reliability separately for ratings of premium, sophisticated, rough, and woodlike.
Reliability was generally high for all rating scales (.76, .81, .75, and .93, respectively), so we
averaged across participants and used the average ratings as predictors for each texture.
A multiple linear regression revealed that premium was the only significant predictor (Full
Model: F(6,40) ¼ 12.2, p < .001, multiple R2 ¼ .65; premium: t(40) ¼ 7.56, p < .001).
Interestingly though, how premium a texture looks accounted for 62.3% of the variance in
texture-preference ratings (r ¼ .79, p < .001), almost identical to the variance accounted for by
16 Perception 0(0)

Figure 5. For a set of 47 visual textures, shows the correlation (r ¼.79, p < .001) between measured
preferences and measured premiumness, averaged across participants. Numbers represent ratings on a line-
mark rating scale, where 100 indicates that the texture is liked ‘‘not at all’’ or is ‘‘not premium,’’ and þ100
indicates that it is liked ‘‘very much’’ or is ‘‘premium.’’

the ecological valence model. Furthermore, premium seemed to track better with preference
than the ecological predictions because it was not compressed relative to preference, nor was
it inflated in the positive direction (see Figure 5). We hypothesize that this is because
participants in our experiments viewed Dislike–Like as synonymous with NotPremium–
Premium.
We should also note that this is not necessarily a critique on Thumfart et al.’s model
because they did not intend the model to be tested in this way. In personal communication, S.
Thumfart indicated that they intended only to investigate the feature set as a whole (i.e., all
188 computational features þ 27 aesthetic antonym pairs) and that any particular prediction
equation from their paper may not be applicable to other samples of textures.
Finally, we did not explicitly test Liu et al.’s instantiation of the model, but we note that
they found the aesthetic property Mussy–Harmonious to be an important predictor of
preference, and we hypothesize that, as with NotPremium–Premium, Mussy–Harmonious
may serve as a synonym for Dislike–Like. Indeed, there is evidence that this may be the case,
at least for colors, because ‘‘preference’’ and ‘‘harmony’’ are often conflated in color research
(for discussion, see Palmer et al., 2013).

General Discussion
The goal of the present study was to test whether the ecological valence theory—already used
to explain color preferences (Palmer & Schloss, 2010) and odor preferences (Schloss,
Goldberger, et al., 2015) —could be extend to account for people’s preferences for visual
Stephens and Hoffman 17

textures. The theory posits that preference for a given stimulus is determined by the combined
valence of all objects associated with that stimulus and that those preferences steer organisms
to approach beneficial outcomes and avoid harmful ones (Palmer & Schloss, 2010; Schloss,
Goldberger, et al., 2015). We found support for the theory for visual textures: The combined
valence of all objects associated with a set of visual textures accounted for 62.5% of the
variance in texture-preference ratings, which was more than was explained by the valence of
the single most frequently associated object for each texture or by the valence of the single
namesake object that produced each texture.
While this prediction accuracy is reasonable, the model did better explaining average
preferences for colors (where it accounted for 80% of the variance; Palmer & Schloss,
2010) and odors (where it accounted for 76% of the variance; Schloss, Goldberger, et al.,
2015). This may seem curious, especially when compared with the model’s performance for
colors: Looking at the visual textures in Figure 2, one might reason that it should be more
obvious what objects are associated with these textures than with a set of uniformly colored
images, so the model should perform better for visual textures than for colors. However, the
objects associated with each visual texture were not as obvious as one would think: Each
texture had an average of 18 descriptions associated with it. Furthermore, visual textures are
not homogeneous fields like colors, and other factors such as composition, style, or context
likely play a role in determining preferences for visual textures (for review, see Palmer et al.,
2013).
The prediction accuracy of the ecological valence model also pales in comparison with the
computational models of Thumfart et al. (2011) and Liu et al. (2015)—the latter of which
claims to account for as much as 99% of the variance in texture-preference ratings—although
we found that the ecological valence model performs similarly to Thumfart et al.’s model
when one uses an optimization measure that emphasizes lower complexity (the ecological
model has only one predictor; Thumfart et al.’s model has six). One problem with the
computational models is their use of some aesthetic properties that participants may
interpret as synonymous with preference (e.g., NotPremium–Premium for Thumfart et al.,
Mussy–Harmonious for Liu et al.). Still, for practical applications (e.g., in marketing or
advertising), a computational model like Thumfart et al.’s or Liu et al.’s is preferable not
only because of the increase in prediction accuracy but also because the end result of such a
model would be able to predict texture preference for novel images based only on
computational features measured from the images (i.e., with no human intervention).
While the ecological valence model has low complexity in a statistical sense—it only uses
one predictor—it has high complexity in a pragmatic sense—computing that single predictor
requires a lot of experimentation.
The real benefit of the EVT is as a theoretical tool. The computational model of Liu et al.
(2015) may be able to account for 99% of the variance in human texture preference, but it
uses predictors that have no theoretical significance. For example, three of the predictors in
Liu et al.’s equation for texture preference come from a set of 20 features that were generated
from a larger set of 106 computational features whose dimensionality was reduced using
stochastic neighbor embedding. Because of the nature of this reduction technique, these
features don’t even have a clear physical interpretation (unlike the starting 106
computational features) let alone a compelling theoretical one. The EVT, on the other
hand, stems from a basic, evolutionarily inspired premise: An observer’s preferences for
low-level stimuli should be driven by the real-world objects those stimuli are most
associated with, and how harmful or beneficial those objects are to the observer. This
premise leads to an entire theory of preferences (i.e., the EVT) and generates testable
predictions.
18 Perception 0(0)

The main limitation of the present study was the range of visual textures used. We used
only 63 visual textures (only 47 for prediction), most of which were Brodatz textures, and all
of which were naturalistic and grayscale. This set of visual textures does not span the range of
possibility, and it could be the case that the ecological valence model performs poorly for
abstract, nonnatural textures. Furthermore, it is known that adding color to a visual texture
affects a person’s emotional response to that texture (e.g., Lucassen, Gevers, & Gijsenij,
2011). Any future study should use a more complete set of textures, perhaps sampling
from a range of texture databases (as in, e.g., Thumfart, Heidl, Scharinger, & Eitzinger,
2009).
In sum, we conclude that—despite the limitations and although there are likely other
mechanisms at play—it is reasonable to think of a person’s preference for a visual texture
as a summary statistic of how they feel about all the objects associated with that texture.

Acknowledgments
We thank Stefan Thumfart and Edwin Lughofer for sharing information about their computational
models. We thank Marcus T. Cole for his help as an independent rater. We thank Charlie Chubb, Greg
Hickok, Daniel Stehr, Darren Peshek, Justin Mark, and Brian Marion for helpful discussions and
feedback. We also thank Steve Palmer and an anonymous reviewer for thoughtful comments and
constructive criticism. Finally, we thank Faizan Memon, Tatiana Dannenbaum, Jonathan Kramer-
Feldman, Angela Babadjanian, and Lilit Barsegyan for assistance running participants.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or
publication of this article.

Funding
The author(s) received no financial support for the research, authorship, and/or publication of this
article.

Notes
1. More formally, if we think of an image as a two-dimensional array of pixels and we think of each
pixel as a realization of a random variable on intensity, then visual textures are those images that
have variation in pixel intensity at the limit of resolution but whose spatial covariance is relatively
homogenous (i.e., moving a small window around the image does not significantly alter the statistics
within the window; Haindl & Filip, 2013; Julesz, 1962; Portilla & Simoncelli, 2000), although other
constraints are often employed and precise definitions vary by application (Haindl & Filip, 2013;
Sebe & Lew, 2001; Tuceryan & Jain, 1998). Also note that visual texture can arise from variations
that are not tactile (e.g., a birch table looks different from a mahogany table, but they might feel
identical).
2. ‘‘Aesthetic properties’’ are operationalized as antonym pairs such as Warm–Cold, Hard–Soft, or
Artificial–Natural. Thus, preference is defined by the antonym pair Dislike–Like.
3. Because of the hierarchical feed-forward nature of these models, preference predictions could
eventually be reduced to only computational features. For example, if ratings of Cold–Warm
turned out to be important for predicting preference, then, in future iterations of the model,
human ratings of Cold–Warm could be replaced by measurements of the computational features
that best predict ratings of Cold–Warm (see Figure 1).
Stephens and Hoffman 19

4. Palmer and Schloss (2010) and Schloss, Goldberger, et al. (2015) refer to their preference predictions
as weighted affective valence estimates, or WAVEs.
5. Recall that we split the participants in the texture-description match rating task into two groups:
Both groups saw all pairings of textures and descriptions ascribed to those textures, but one group
saw the same description for all textures given that description before moving on to the next
description, and the other group saw the same texture for all descriptions given to that texture
before moving on to the next texture. The latter group went significantly slower (mean difference
in reaction time per trial ¼ 2.05 s, t(24) ¼ 10.36, p < .001), but, as Cronbach’s alpha indicates, the
groups’ ratings were in strong agreement, and using data from either group led to the same
correlation coefficient.
6. Thumfart et al. (2011) actually used a computational feature termed roughness by Tamura et al.
(1978). However, Tamura et al. designed the feature roughness to match human ratings of Rough–
Smooth as closely as possible. Furthermore, we found that using the feature roughness in lieu of
ratings of Rough–Smooth did not change our results.
7. When reading a preliminary version of this article, Charlie Chubb pointed out that the namesake
hypothesis still explains a good amount of the variance considering it has a significant additional
constraint. The problem with comparing the namesake hypothesis model to the full EVT model is
that both models lack parameters so there’s no way to quantify this additional constraint (i.e., the
models can’t be compared with a traditional methodology such as a likelihood ratio). This problem
also arises when comparing the single-associate hypothesis model to the full EVT model.

References
Amadasun, M., & King, R. (1989). Textural features corresponding to textural properties. IEEE
Transactions on Systems, Man and Cybernetics, 19, 1264–1274.
Anderson, B. L., & Kim, J. (2009). Image statistics do not explain the perception of gloss and lightness.
Journal of Vision, 9, 10.1–17.
Ben-Shahar, O. (2006). Visual saliency and texture segregation without feature gradient. Proceedings of
the National Academy of Sciences of the USA, 103, 15704–15709.
Bergen, J. R. (1991). Thoeries of visual texture perception. In D. Regan (Ed.), Vision and visual
dysfunction (Vol. 10B, pp. 114–134). New York, NY: Macmillan.
Brainard, D. H. (1997). The psychophysics toolbox. Spatial Vision, 10, 433–436.
Brodatz, P. (1966). Textures: A photographic album for artists and designers. New York, NY: Dover.
Chen, Y., Sobue, S., & Huang, X. (2008). Mapping functions of color image features and human
KANSEI. Paper presented at IIHMSP ’08 International Conference on Intelligent Information
Hiding and Multimedia Signal Processing (pp. 725–728). Harbin, China: IEEE.
Cronbach, L. J. (1951). Coefficient alpha and the internal structure of tests. Psychometrika, 16, 197–334.
Dong, Y., Tao, D., Li, X., Ma, J., & Pu, J. (2015). Texture classification and retrieval using shearlets
and linear regression. IEEE Transactions on Cybernetics, 45, 358–369.
Gatto, J. A., Porter, W. A., & Selleck, J. (2010). Exploring visual design: The elements and principles
4th ed. Worcester, MA: Davis Publications.
Guo, Z., Zhang, L., & Zhang, D. (2010). Rotation invariant texture classification using LBP variance
(LBPV) with global matching. Pattern Recognition, 43, 706–719.
Haindl, M., & Filip, J. (2013). Visual texture: Accurate material appearance measurement, representation
and modeling. London, England: Springer-Verlag.
Haralik, R. M., Shanmugam, K., & Dinstein, I. (1973). Texture features for image classification. IEEE
Transactions on Systems, Man and Cybernetics, 3, 610–621.
Ho, Y. X., Landy, M. S., & Maloney, L. T. (2006). How direction of illumination affects visually
perceived surface roughness. Journal of Vision, 6, 634–648.
Joshi, D., Datta, R., Fedorovskaya, E., Luong, Q., Wang, J. Z., Li, J., . . . Luo, J. (2011). Aesthetics and
emotions in images. IEEE Signal Processing Magazine, 28, 94–115.
Julesz, B. (1962). Visual pattern discrimination. IEEE Transactions on Information Theory, 8,
84–92.
20 Perception 0(0)

Julesz, B. (1981). Textons, the elements of texture perception, and their interactions. Nature, 290,
91–97.
Kawamoto, N., & Soen, T. (1993). Objective evaluation of color design II. Color Research and
Application, 18, 260–266.
Kim, J., & Anderson, B. L. (2010). Image statistics and the perception of surface gloss and lightness.
Journal of Vision, 10, 3.
Kim, S., Kim, E. Y., Jeong, K., & Kim, J. (2006). Emotion-based textile indexing using colors, texture
and patterns. In G. Bebis, et al. (Eds.), Advances in visual computing (pp. 9–18). Berlin, Germany:
Springer.
Kleiner, M., Brainard, D., & Pelli, D. (2007). What’s new in Psychtoolbox-3? Perception, 36, 14. (ECVP
Abstract Supplement).
Landy, M. S., & Bergen, J. R. (1991). Texture segregation and orientation gradient. Vision Research, 31,
679–691.
Landy, M. S., & Graham, N. (2004). Visual perception of texture. In L. M. Chalupa, & J. S. Werner
(Eds.), The visual neurosciences (pp. 1107–1118). Cambridge, MA: MIT Press.
Lee, J., & Park, E. (2011). Fuzzy similarity-based emotional classification of color images. IEEE
Transactions on Multimedia, 13, 1031–1039.
Liu, J., Lughofer, E., & Zeng, X. (2015). Could linear model bridge the gap between low-level statistical
features and aesthetic emotions of visual textures? Neurocomputing, 168, 947–960.
Lucassen, M. P., Gevers, T., & Gijsenij, A. (2011). Texture affects color emotion. Color Research and
Application, 36, 426–436.
Malik, J., & Perona, P. (1990). Preattentive texture discrimination with early vision mechanisms.
Journal of the Optical Society of America A, 7, 923–932.
Montoyishi, I., Nishida, S., Sharan, L., & Adelson, E. H. (2007). Image statistics and the perception of
surface qualities. Nature, 447, 206–209.
Palmer, S. E., & Schloss, K. B. (2010). An ecological valence theory of human color preference.
Proceedings of the National Academy of Sciences of the USA, 107, 8877–8882.
Palmer, S. E., Schloss, K. B., & Sammartino, J. (2013). Visual aesthetics and human preference. Annual
Review of Psychology, 64, 77–107.
Portilla, J., & Simoncelli, E. (2000). A parametric texture model based on joint statistics of complex
wavelent coefficients. International Journal of Computer Vision, 40, 49–71.
Randen, T., & Husy, J. H. (1999). Filtering for texture classification: A comparative study. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 21, 291–310.
Rao, A. R., & Lohse, G. (1993). Towards a texture naming system: Identifying relevant dimensions of
texture perception. (Research Report No RC19140 (83352)). Almaden: IBM Research Division.
Rao, A. R., & Lohse, G. L. (1996). Towards a texture naming system: Identifying relevant dimensions
of texture. Vision Research, 36, 1649–1669.
Ritter, N. L. (2010). Understanding a widely misunderstood statistic: Cronbach’s alpha. Paper presented
at Southwestern Educational Research Association Conference, New Orleans, LA.
Rosenholtz, R. (2000). Significantly different textures: A computational model of pre-attentive texture
segmentation. In V. David (Ed.), Computer vision – ECCV 2000 (pp. 197–211). Berlin, Germany:
Springer.
Rosenholtz, R. (2015). Texture perception. In J. Wagemans (Ed.), Oxford handbook of perceptual
organization (pp. 167–186). Oxford, England: Oxford University Press.
Schloss, K. B., Goldberger, C. S., Palmer, S. E., & Levitan, C. A. (2015). What’s that smell? An
ecological approach to understanding preferences for familiar odors. Perception, 44, 23–38.
Schloss, K. B., Hawthorne-Madell, D., & Palmer, S. E. (2015). Ecological influences on individual
differences in color preference. Attention, Perception, and Psychophysics, 77, 2803–2816.
Schloss, K. B., & Palmer, S. E. (2014). The politics of color: Preference for republican-red vs.
democtratic-blue. Psychonomic Bulletin and Reivew, 21, 1481–1488.
Schloss, K. B., Poggesi, R. M., & Palmer, S. E. (2011). Effects of university affiliation and ‘‘school spirit’’
on color preferences: Berkeley versus Stanford. Psychonomic Bulletin and Review, 18, 498–504.
Stephens and Hoffman 21

Sebe, N., & Lew, M. S. (2001). Texture features for content-based retrieval. In M. S. Lew (Eds.),
Principles of visual information retrieval (pp. 51–85). London, England: Springer-Verlag.
Tamura, H., Mori, S., & Yamawaki, T. (1978). Textural features corresponding to visual perception.
IEEE Transactions on Systems, Man and Cybernetics, 8, 460–473.
Taylor, C., & Franklin, A. (2012). The relationship between color-object associations and color
preference: Further investigation of the ecological valence theory. Psychonomic Bulletin and
Review, 19, 190–197.
Thumfart, S., Heidl, W., Scharinger, J., & Eitzinger, C. (2009). A quantitative evaluation of texture
feature robustness and interpolation behaviour. In X. Jaing, & N. Petkov (Eds.), Computer analysis
of images and patters: Lecture notes in computer science (Vol. 5702, pp. 1154–1161). Berlin, Germany:
Springer.
Thumfart, S., Jacobs, R. H., Lughofer, E., Eitzinger, C., Cornelissen, F. W., Groissboeck,
W., . . . Richter, R. (2011). Modeling human aesthetic perception of visual textures. ACM
Transactions on Applied Perception, 8, Article 27.
Tuceryan, M., & Jain, A. (1998). Texture analysis. In C. H. Chen, L. F. Pau, & P. S. Wang (Eds.), The
handbook of pattern recognition and computer vision (2nd ed. pp. 207–248). Hackensack, NJ: World
Scientific Publishing Co.
Tyler, C. W. (2004). Theory of texture discrimination of based on higher-order perturbations in
individual texture samples. Vision Research, 44, 2179–2186.
Um, J., Eum, K., & Lee, J. (2002). A study of the emotional evaluation models of color patterns based
on the adaptive fuzzy system and the neural network. Color Research and Application, 27, 208–216.
Varma, M., & Zisserman, A. (2005). A statistical approach to texture classification from single images.
International Journal of Computer Vision, 62, 61–81.
Victor, J. D. (1988). Models for preattentive texture discrimination: Fourier analysis and local feature
processing in a unified framework. Spatial Vision, 3, 263–280.
Yokosawa, K., Yano, N., Schloss, K. B., Prado-Leon, L. R., & Palmer, S. E. (2010). Cross-culteral
studies of color preferences: US, Japan, and Mexico [abstract]. Journal of Vision, 10, 408.
Zhang, H., Augilius, E., Honkela, T., Laaksonen, J., Gamper, H., & Alene, H. (2011). Analyzing
emotional semantics of abstract art using low-level image features. In J. Gama, E. Bradley, & J.
Hollmén (Eds.), Proceedings of the 10th International Conference on Advances in Intelligent Data
Analysis, 413–423.

Appendix A
Table A1 shows all of the information needed to run the calculations in this article for the 47
visual textures that we predicted preferences for. Table A2 shows information for the
remaining 15 textures whose preferences we did not predict. The title (first column for
both tables) of each texture is either the description given by Brodatz (1966), with his
numbering in parentheses, or the search term that produced the image on Shutterstock.

Appendix B
Equation B1 gives the details for Equation 2 in the article and is adapted from Thumfart et al.
(2011):
MAE Model Complexity
MAEpunished ¼  þ ð1  Þ  ðB1Þ
rangeðMAEÞ maxðModel ComplexityÞ

where MAE stands for ‘‘mean absolute error.’’ For a given visual texture, the absolute error
is the absolute value of the difference between the measured and predicted preferences for
that texture. Averaging the absolute error across all textures gives the MAE, and subtracting
 22

Table A1. Data for the 47 Visual Textures Whose Preference Was Predicted.

Average EVT Average Most frequent Average Namesake Average

preference predicted Number of premium description Valence description Valence
Visual texture rating preference descriptions rating (MFD) of MFD (ND) of ND

Wire mesh (D1) 2.46 2.05 17 14.77 Fence/gate 8.67 Wire mesh 13.33
Crocodile skin (D10) 27.65 1.10 29 7.98 Reptile skin (snake, 25.66 Reptile skin (snake, 25.66
crocodile, lizard) crocodile, lizard)
Japanese rice paper 11.96 3.99 24 31.49 Branches/roots (of a 15.86 Paper 25.66
(D107) tree)
Grassy fiber (D110) 9.16 6.44 16 25.24 Grass 31.57 Grass 31.57
Plastic bubbles (D112) 16.52 5.31 24 5.72 Beehive 29.03 Bubbles 35.6
Straw (D15) 10.7 9.24 14 34.35 Grass 31.57 Straw 10.94
Raffia weave (D18) 21.64 11.44 21 15.65 Woven basket 20.31 Woven basket 20.31
Beach pebbles (D23) 31.31 20.13 12 7.75 Rocks/pebbles/stones 22.35 Rocks/pebbles/stones 22.35
Brick wall (D26) 19.21 6.28 6 6.41 Bricks (brick wall/brick 4.42 Bricks (brick wall/brick 4.42
building/brick building/brick
sidewalk) sidewalk)
Reptile skin (D3) 26.96 1.72 25 3.99 Reptile skin (snake, 25.66 Reptile skin (snake, 25.66
crocodile, lizard) crocodile, lizard)
Beach pebbles (D30) 32.32 12.90 13 21.21 Rocks/pebbles/stones 22.35 Rocks/pebbles/stones 22.35
Pressed cork (D32) 2.76 4.43 22 28.99 Wall of a house/ 13.02 Pressed cork –
building
Netting (D34) 6 2.07 21 6.64 Net/netting 6.08 Net/netting 6.08
Water (D37) 26.7 24.46 9 14.85 Ocean/sea 67.34 Ocean/sea 67.34
Lace (D41) 38.29 28.34 17 35.32 Flowers 59.28 Lace/embroidery 35.15
Mica (D5) 6.13 13.64 14 17.51 Rocks/pebbles/stones 22.35 Wall of a house/ 13.02
building
Raffia woven with 14.74 11.55 12 10.07 Tree bark 12.47 Cloth/fabric 27.49
cotton (D50)
Oriental straw cloth 6.85 7.61 19 17.04 Cloth/fabric 27.49 Cloth/fabric 27.49
(D52)
Handmade paper 10.11 8.64 20 29.71 Wall of a house/ 13.02 Paper 25.66
(D57) building
Perception 0(0)

(continued)
Table A1. Continued.

Average EVT Average Most frequent Average Namesake Average

preference predicted Number of premium description Valence description Valence
Visual texture rating preference descriptions rating (MFD) of MFD (ND) of ND

European marble 12.87 0.12 26 28.28 Ocean/sea 67.34 Granite/marble 31.39

(D58)
Stephens and Hoffman

European marble 15.84 4.76 24 8.30 Rocks/pebbles/stones 22.35 Granite/marble 31.39

(D61)
European marble 0.62 14.52 14 0.89 Rocks/pebbles/stones 22.35 Granite/marble 31.39
(D62)
Handwoven rattan 30.31 10.30 16 8.39 Woven basket 20.31 Woven basket 20.31
(D64)
Wood grain (D69) 19.19 13.40 11 28.18 Tree bark 12.47 Wood 28.52
Wood grain (D71) 22.48 10.48 18 24.25 Wood 28.52 Wood 28.52
Tree stump (D72) 19.39 13.01 6 4.79 Tree bark 12.47 Tree bark 12.47
Soap bubbles (D73) 16.19 4.18 34 32.20 Bubbles 35.6 Bubbles 35.6
Coffee beans (D74) 13.69 13.04 12 17.61 Coffee beans 40.37 Coffee beans 40.37
Oriental straw cloth 10.15 14.26 18 21.13 Wood 28.52 Cloth/fabric 27.49
(D78)
Ceiling tile (D86) 17.99 5.65 22 39.24 Wall of a house/ 13.02 Wall of a house/ 13.02
building building
Fossilized sea fan with 2.99 6.94 23 4.01 Leaves 34.87 Coral/coral reef 42.84
coral (D87)
Dried hop flowers 10.15 8.11 20 1.24 Insects/bugs 35.54 Flowers 59.28
(D88)
Grass lawn (D9) 21.42 2.72 21 41.91 Grass 31.57 Grass 31.57
Clouds (D90) 0.83 3.49 16 6.46 Clouds 61.82 Clouds 61.82
Clouds (D91) 63.88 22.74 8 27.75 Clouds 61.82 Clouds 61.82
Fur, hide of unborn calf 10.78 5.47 6 5.47 Animal fur (e.g., dog, 6.69 Animal fur (e.g., dog, 6.69
(D93) cat) cat)
Crushed rose quartz 18.11 12.65 16 8.60 Rocks/pebbles/stones 22.35 Crystals/diamonds/ 52.46
(D98) gems
Chocolate spread 28.67 10.07 20 38.47 Chocolate 61.7 Chocolate 61.7
(continued)
23
 24

Table A1. Continued.

Average EVT Average Most frequent Average Namesake Average

preference predicted Number of premium description Valence description Valence
Visual texture rating preference descriptions rating (MFD) of MFD (ND) of ND

Chocolate (Gaussian 54.01 4.17 21 43.37 X-ray/ultrasound 1.04 – –

blurred)
Mud 51.91 1.52 23 49.88 Mud 21.09 Mud 21.09
Mud (Gaussian 66.27 1.64 21 63.15 People 40.24 – –
blurred)
Lettuce 49.81 22.30 15 34.32 Flowers 59.28 Lettuce 37.99
Mold 23.27 1.50 29 41.54 Dirt/gravel/soil 9.74 Mold 66.16
Rotted strawberry 48.37 0.27 27 25.08 Beans 8.09 Rotten/spoiled food 74.77
Strawberry 12.57 9.31 27 1.84 Eggs 33.25 Strawberry 49.51
Infected skin 26.48 0.78 28 17.61 Ocean/sea 67.34 Infection 71.47
Tropical sea 36.73 31.83 9 30.20 Ocean/sea 67.34 Ocean/sea 67.34

Note. All rating scales (i.e., preference, premium, and valence) vary from 100 to þ100. There is no valence rating for pressed cork because it was not listed during the object-
description task, and the blurred chocolate and mud were not given namesake descriptions.
Perception 0(0)
Stephens and Hoffman 25

Table A2. Data for the 15 Visual Textures Whose Preference Was Not Predicted.

Average Average Average

preference Number of premium Most frequent Valence
Visual texture rating descriptions rating description (MFD) of MFD

Woven cane (D101) 1.88 19 15.22 Cloth/fabric 27.49

Woolen cloth (D11) 12.13 17 6.75 Clothing (shirt/pants/ 57.07
sweater/jacket)
Brick wall (D25) 18.43 6 11.12 Bricks (brick wall/brick 4.42
building/brick
sidewalk)
Beach sand (D29) 5.24 17 29.34 Cement/concrete/ 7.43
pavement
Beach pebbles (D31) 45.84 10 19.52 Rocks/pebbles/stones 22.35
Lace (D39) 11.18 28 7.95 Flowers 59.28
Lace (D40) 38.27 23 42.47 Flowers 59.28
Lace (D42) 35.24 23 44.54 Curtain 20.1
Straw matting (D55) 2.11 20 1.38 Clothing (shirt/pants/ 57.07
sweater/jacket)
European marble (D63) 21.67 31 19.67 Spider web 40.72
Oriental straw cloth (D80) 20.28 21 28.33 Carpet/rug 15.51
Oriental straw cloth (D82) 1.41 24 7.00 Carpet/rug 15.51
Woven matting (D83) 7.91 21 0.58 Cloth/fabric 27.49
Dried hop flowers (D89) 26.21 23 8.20 Insects/bugs 35.54
Brick wall (D94) 31.42 5 3.01 Bricks (brick wall/brick 4.42
building/brick
sidewalk)

Note. All rating scales vary from 100 to þ100.

the maximum absolute error from the minimum gives range(MAE). Thus, MAE and
range(MAE) could easily be computed for the ecological valence model from the data in
the second and third columns of Table A1 in Appendix A. ModelComplexity is given by the
number of predictors/regressors in the model. In Thumfart et al.’s model, they used up to 50
predictors so that max(ModelComplexity) ¼ 50.
Table B1 shows the values for MAE, range(MAE), ModelComplexity, and
max(ModelComplexity) for the ecological valence model versus Thumfart et al.’s
computational model (Note: MAE for the computational model is given in Table A1 of
Thumfart et al., 2011, and range(MAE) was calculated using Equation B1 after E.
Lughofer indicated in personal communication that they used an  value of .90).
Using Equation B1 and the values in Table B1, it’s straightforward to show that
MAEpunished EVT ¼ MAEpunished Thum ¼ 0.10 when  ¼ .35.

Table B1. Values for the Variables in Equation B1.

Model MAE range(MAE) Model Complexity max(Model Complexity)

EVT model 17.29 67.50 1 50

Thumfart et al. model 9.57 143.55 6 50

Note. EVT ¼ecological valence theory; MAE ¼ mean absolute error.