Exp Brain Res (2007) 181:199–211
DOI 10.1007/s00221-007-0923-2

R ES EA R C H A R TI CLE

Scan patterns during the processing of facial identity
in prosopagnosia
Jason J. S. Barton · Nathan RadcliVe ·
Mariya V. Cherkasova · Jay A. Edelman

Received: 4 July 2006 / Accepted: 19 February 2007 / Published online: 15 March 2007
© Springer-Verlag 2007

Abstract The scan patterns of ocular Wxations made by
prosopagnosic patients while they attempt to identify faces
may provide insights into how they process the information
in faces. Contrasts between their scanning of upright versus
inverted faces may index the presence of a hypothesized
orientation-dependent expert mechanism for processing
faces, while contrasts between their scanning of familiar
versus novel faces may index the inXuence of residual
facial memories on their search for meaningful facial information. We recorded the eye movements of two prosopagnosics while they viewed faces. One patient, with acquired
prosopagnosia from a right occipitotemporal lesion,
showed degraded orientation eVects but still with a normal
distribution of Wxations to more salient facial features.

J. J. S. Barton · M. V. Cherkasova
Department of Neurology,
Beth Israel Deaconess Medical Center,
Harvard Medical School, Boston, MA, USA

However, the dynamics of his global scan patterns were
more chaotic for novel faces, suggesting degradation of an
internal facial schema, and consistent with other evidence
of impaired face conWguration perception in this patient.
His global scan patterns for famous faces diVered from
novel faces, suggesting the inXuence of residual facial
memories, as indexed previously by his relatively good
imagery for famous faces. The other patient, with a developmental prosopagnosia, showed anomalous orientation
eVects, abnormal distribution of Wxations to less salient
regions, and chaotic global scan patterns, in keeping with a
more severe loss of face-expert mechanisms. The eVects of
fame on her scanning were weaker than those in the Wrst
subject and non-existent in her global scan patterns. We
conclude that scan patterns in prosopagnosia can both reXect the loss of orientation-dependent expert mechanisms
and index the covert inXuence of residual facial memories.
In these two subjects the scanning data were consistent with
other results from tests of conWguration perception, imagery, and covert recognition.

J. J. S. Barton
Division of Neurology, Department of Ophthalmology
and Visual Sciences, University of British Columbia,
Vancouver, BC, Canada

Keywords Face recognition · Inversion ·
Scanpath · Morph · Eye movement · Prosopagnosia

J. A. Edelman
Department of Psychology, Harvard University,
Cambridge, MA, USA

Introduction

N. RadcliVe
Faculty of Medicine, Temple University,
Philadelphia, PA, USA
J. J. S. Barton (&)
Neuro-ophthalmology Section D,
VGH Eye Care Center, 2550 Willow Street,
V5Z 3N9 Vancouver, BC, Canada
e-mail: jasonbarton@shaw.ca

Inspection of objects or scenes is usually accomplished
with a series of scanning saccades that shift ocular Wxation
from one location to another. It is unlikely that these Wxations are randomly distributed across the image: rather,
their purpose may be to direct both the locus of attention
and the high spatial resolution of the macula to regions that
will yield information of use to the observer. From this
goal-directed stance, scanning acquires data for a cognitive

123

200

process aimed at reaching perceptual decisions about the
world (Rybak et al. 1998; Deco and Schurmann 2000;
Henderson 2003). Studying scan patterns can tell us where
useful information is located in a scene. Furthermore,
because scanning is a quintessentially serial process, scan
patterns can also inform us about the temporal organization
of information processing within the observer.
This concept of scanning as both a manifestation of and
an instrument for information processing is consistent with
observations that scan patterns reXect properties of both the
stimulus and the observer. Attention and Wxations can be
attracted by properties such as color contrast, high-contrast
edges, and relative motion, which may indicate regions in
the scene containing high-yield information. On the other
hand, the goals of the observer may alter scanning: the
same observer will scan a city scene diVerently when looking for an address than when looking for a person in the
crowd. The observer’s prior experience and expertise with
stimuli can also inXuence scan patterns. Art experts scan
pictures diVerently from naïve observers (Zangemeister
et al. 1995), for example, and experimental subjects scan
familiar items diVerently than novel items (AlthoV and
Cohen 1999). These illustrate important ‘top-down’ eVects
of the cognitive status of the individual upon the scan patterns used to acquire information from a scene.
Because of these observer-dependent eVects, it may be
possible to use scan patterns as a probe of the perceptual
capabilities of individuals with or without neurological disorders. Patients with prosopagnosia, for example, have a
speciWc dysfunction that impairs their ability to recognize
familiar faces. In some of these individuals the deWcit may
lie at a perceptual level, degrading their ability to acquire
perceptual data about faces (Barton et al. 2002). This perceptual function is a highly expert process, able to discern
the very subtle changes that distinguish one speciWc face
from the thousands of other faces encountered during a lifetime. If this expert mechanism is lost, it may be possible to
discern its absence in altered scan patterns of prosopagnosic patients viewing faces.
Other literature from normal subjects suggests that one
of the hallmarks of a face-expert processing mechanism is
the inversion eVect. Faces are more diYcult to recognize
upside down (Yin 1970), likely because inversion impairs
the perception of facial structure (Sergent 1984; Kemp
et al. 1990; Searcy and Bartlett 1996; Leder and Bruce
1998; Barton et al. 2001b, 2003b; Malcolm et al. 2004).
The dual-mode hypothesis (Bartlett and Searcy 1993; Farah
et al. 1998) explains this by postulating that faces can be
processed by both generic object recognition and a trained
expert mechanism; however, since expertise is normally
acquired through repeated exposure to upright faces, the
trained expert mechanism develops an orientation dependency, so that it is operative with only upright but not

123

Exp Brain Res (2007) 181:199–211

inverted faces. Our prior work showed that normal subjects
scan upright faces diVerently than inverted faces, with less
overall scanning, less scanning of the less salient lower
face, and a less random sequence of Wxations (Barton et al.
2006). If an orientation-dependent expert mechanism for
looking at faces is disrupted in prosopagnosia, then their
scan patterns may fail to show these diVerences between
upright and inverted viewing of faces.
In addition to showing eVects of expertise, scan patterns
may also show eVects of experience, or familiarity. Wellknown faces possess representations within the subject’s
store of facial memories, and this knowledge may guide the
exploratory scan patterns of subjects as they engage in the
process of identifying the face. Contrasts between famous
and recently encountered (novel) faces have shown that
familiarity generates a variety of eVects on scan patterns
(AlthoV and Cohen 1999; Stacey et al. 2005; Barton et al.
2006). In prosopagnosia, such scanning diVerences
between famous and novel faces could index the residual
eVects of surviving stores of facial memories, even if the
patients were not able to recognize the faces being shown.
Such a dissociation between absent overt recognition and
residual behavioral eVects that distinguish between familiar
and novel faces could also be considered a form of covert
recognition in this disorder (Bruyer 1991; Young 1994).
Indeed there is one previous report of such diVerences in
the face scan patterns in two prosopagnosic patients (Rizzo
et al. 1987).
In this study we had subjects engage in a face identiWcation task while we manipulated the variables of expertise
(upright versus inverted) and familiarity (famous versus
novel). We used a series of morphed faces to increase the
level of diYculty in the identiWcation task for control subjects, and hence encourage the generation of scan patterns
during a more protracted period of information processing.
We studied two prosopagnosic patients, one with an
acquired lesion and one with the developmental form. Our
goal was to determine if either subject demonstrated residual expertise in the perceptual processing of faces, or residual familiarity eVects suggesting the presence of facial
memories and a form of covert recognition.

Methods
Subjects
Two prosopagnosic patients participated in the study.
Subject 002(D) is a 31-year-old woman with generalized
seizures since 2 years of age. She has always had diYculty
recognizing faces, using cues such as voice or clothing
instead. She also has trouble recognizing her brother’s car
in parking lots. She gets lost in familiar surroundings and

Exp Brain Res (2007) 181:199–211

her reading is slow, but she denied problems with color
perception. Her visual acuity was 20/50. She saw 11 of 14
Ishihara pseudoisochromatic plates, in the normal range.
Confrontation visual Welds were only mildly constricted in
the periphery. She had left superior oblique overaction but
otherwise normal eye movements. Neurologic examination
was normal. Neuropsychologic testing showed average verbal intelligence and superior attention and working memory. Memory for verbal material was normal, but that for
non-verbal material was severely impaired. Object recognition was impaired, for unusual views and line drawings.
Reading speed and comprehension were slow. There was
no evidence of hemi-neglect. On a test of discrimination
between familiar and unfamiliar faces, she scored a d’ of
zero, indicating no discriminative ability. Forced-choice
tests also showed no evidence of covert recognition (Barton
et al. 2001a). Her perception of facial conWguration was
severely impaired (Barton et al. 2003a). MRI showed
abnormal gyral patterns of both occipital lobes, interpreted
as a mild focal developmental polymicrogyria (Fig. 1, top).
Subject 005(A) is a 52-year-old man seen 7 months after a
right occipital hemorrhage from an underlying oligodendroglioma. This was resected and he was treated with irradiation
and chemotherapy. He complained of blurry vision and diYculty reading. He had trouble recognizing buildings in his
hometown, and got lost occasionally. Snellen acuity was
20/25 in either eye, and he had a complete left homonymous
hemianopia on Goldmann perimetry. He saw 12/14 Ishihara
pseudoisochromatic plates. Pursuit and saccadic eye movements were normal. There were no signs of left hemi-neglect

201

or simultanagnosia. On the test of famous face familiarity, he
scored a d’ of 1.12, indicating very weak overt recognition of
faces (Barton et al. 2001a). His perception of facial conWguration was severely impaired (Barton et al. 2002), but his
imagery for faces was normal, particularly for facial features,
suggesting that facial memory stores were intact (Barton and
Cherkasova 2003). He had evidence of covert recognition on
both forced-choice tests (Barton et al. 2001a) and an indirect
priming method (Barton et al. 2004). MRI showed an extensive right medial occipital resection involving the lingual and
fusiform gyri (Fig. 1, bottom).
As controls, we studied eight normal subjects, three
female and Wve male, with a mean age of 31.5 years (range
23–41 years). All subjects gave informed consent and the
institutional review board of Beth Israel Deaconess Medical Center approved the protocols.
Images
We created a set of eight morph series, where each series
represented a gradual physical transformation from one
face to a second. Four series were transitions between two
male faces, and four were transitions between two female
faces. In each gender set, two series used pairs of famous
faces and two used pairs of novel faces.
For famous faces, we selected eight images with a highresolution frontal view from the internet. We rejected
images with excess jewelry, make-up, facial hair, open
mouths, or glasses. All subjects, both prosopagnosic and
control, acknowledged having heard of and having had the

Fig. 1 MR scans of patients.
Top Wgure is an axial
T2-weighted image from subject
002(D), showing the
polymicrogyria in the medial
occipital cortex (arrows).
Bottom Wgures are coronal
T1-weighted images from subject 005(A), with gadolinium
enhancement in the right image.
These show the medial
occipitotemporal aspects of the
lesion, involving the general
vicinity of the fusiform face area
(arrows)

123

202

opportunity to see the faces of these celebrities. In particular, the two patients were able to provide semantic information about the celebrities in response to their name, even
though they were not able to identify their faces or recognize the faces as being familiar. For novel faces, we used a
digital camera to take pictures of eight anonymous individuals, ensuring that their images matched the above criteria
for famous faces.
Using Photoshop 5.5 (Adobe, San Jose, CA, USA), we
replaced the backgrounds with a black surround, removed
earrings, and cropped the images to eliminate the external
hair contour and neck. Morph pairs were matched for hair
color, facial shape, image size, image luminance, sex, and
fame. The celebrity pairs were Demi Moore/Julia Roberts,
Michelle PfeiVer/Gwyneth Paltrow, John Travolta/Paul
Newman, Matthew Broderick/Kevin Costner.
For each morph series the two base (unmorphed) images
were loaded into Morph Studio 1.0, where 120 points were
matched between each image on analogous facial regions
(i.e., center of pupils, corner of mouth). We created a
morph series of 21 images, representing a 5% change in
position, hue and contrast between the two base images
(designated images 1 and 21). From each series we used the
two base images (images 1 and 21, representing mixtures of
{0:100}) and Wve morphed images from near the center of
each series (images 9, 10, 11, 12, and 13, representing mixtures of {40:60}, {45:55}, and {50:50}).
Each image was set to a height of 330 pixels, and
because face shapes varied, widths ranged from 299 pixels
to 392 pixels. These generated relatively large images, which
improve the accuracy of locating Wxations to a given region
of interest. Cropping of hair was performed to enforce
attention to facial details. It is possible that the use of large
images, omission of hair contour, and the use of morphed
images limit the ability to compare these data to other studies, but we believe these maneuvers enhanced the accuracy
of the results and focus the eye movements upon the face
processing task at hand.
Eye movement protocol
We recorded eye movements with a magnetic search coil
technique, using a scleral contact lens and a three-foot Weld
coil (Crist Instruments, Bethesda, MD, USA). The subjects’
heads were secured in a chin rest. The lens was placed in
the left eye. We calibrated the system by having subjects
Wxate nine targets in a square grid spanning 50°. Twelve
data points were collected at each of nine targets, and a
regression method was used to Wnd the best linear Wt. Eye
position was digitized at 500 samples/s.
Displays were generated by a Power Macintosh 9600/
233 using programs written in C++ on the Vision Shell
programming platform (www.kagi.com/visionshell), and

123

Exp Brain Res (2007) 181:199–211

back-projected onto the tangent screen with an Eiki LC7000U projector. Distance from the subject’s cornea to the
screen was 80.5 cm. On the screen, each facial image
spanned a height of 67° with widths ranging from 61° to 82°.
We divided our faces into two experimental sets, one
male and one female. Before recordings, we showed the
subjects four panels, each containing an array of eight
faces, all female or all male. Two of these arrays showed
the base images used for the morphed series, while the
other two showed alternate pictures of the same individuals.
The faces had the names of the individuals printed below
them, and subjects were asked to try to remember the identity of the faces. They were allowed to study these panels as
long as they wished. They were told that in the experiment
they would be seeing a sequence of faces that had been
altered, and that they would be asked to identify to whom
each face belonged.
We presented our faces in two blocks, one female and
one male. The order of male/female blocks was counterbalanced across the eight subjects. Within each block, half the
trials were upright and half inverted. In each block, morph
series (e.g., Demi Moore/Julia Roberts), morph image (e.g.,
40Demi:60Julia), and orientation were randomly ordered.
Each face image was seen just twice, once in each orientation. Thus each subject saw a total of 112 faces (seven
morph levels, four same-sex face pairs, two orientations,
two diVerent-sex blocks).
A trial began with a Wxation screen with a small red spot
at center. Fixation within 3° of this zone triggered presentation of a stimulus face, along with a red spot 30° below
screen center. Subjects were asked to look at this second red
spot when they had reached a decision, then state whether
this face was famous or novel. On their report the collection
of eye movement data stopped and the single face was
replaced with an array of the eight alternate upright facial
images–(i.e., the faces not used as base images in the morphed series. These alternates were used to emphasize the
identity recognition element of the task, rather than pattern
matching). Each of the eight faces had a number below it
and the subject indicated the number of the individual whose
face they thought they had just seen. When they had done
so, the panel vanished, the Wxation screen for the next trial
appeared, and collection of eye movement data resumed.
After every ten trials a calibration screen appeared, with
Wve Wxation spots, one at center and one at the four corners
of the region where the faces were displayed. These were
numbered, and subjects were asked to look at each spot in
order, and then to return to screen center.
Data analysis
Using Matlab (Mathworks, Natick, MA, USA), we identiWed saccades as eye movements with velocities exceeding

Exp Brain Res (2007) 181:199–211

47°/s. The onset of a saccade was the point at which the
velocity of the eye Wrst exceeded 31°/s, and the end of a
saccade as a point where velocity fell below this baseline.
Fixations were deWned as the intervals between the end of
one saccade and the beginning of the next saccade. We
characterized each Wxation by its horizontal position, vertical position, duration, and its time of onset relative to the
start of the trial.
We recorded all Wxations from the beginning to end of
each trial. The initial Wxation, evident as a long Wxation of
several seconds at screen center, was discarded, and the
Wrst Wxation after this long Wxation was taken as the onset
of the scanning sequence. The end of the scanning
sequence was the point at which subjects Wxated the red
dot at 30° below screen center. Not all trials were terminated by such a Wxation, however, as subjects often forgot
to Wxate this peripheral spot when they had made a familiarity decision. In these trials, a single long Wxation of
more than 1,500 ms duration was often present near the
end of the string of eye movement data, usually at an
eccentric position in relation to the image, and this was
taken to indicate the end of the subject’s scanning. In the
absence of either a Wxation at the red spot or a long nearterminal Wxation, we included all Wxations occurring to the
end of data collection for a trial.
Our next step was to adjust our data for any baseline
drift or gain change that might have occurred during
recording, since absolute eye position was critical to our
results. From the recurring calibration trials we obtained the
average eye position for Wxations near each of the Wve
spots. OVset shift was calculated as the mean diVerence in x
and y positions from the expected value for each point.
Gain shift was calculated as the mean diVerence in distance
of each of the four corner Wxations from the Wxations at
screen center (gain shift proved to be negligible for all data
sets). We derived the oVset and gain correction required by
averaging these calculations from the two calibration trials
sandwiching a set of ten face trials, and applied to all ten
trials. We then collated the corrected data from all subjects
for each trial, displayed these data on an x–y plot of position upon the facial image of that trial, and identiWed all
Wxations that fell upon the face.
Our analysis of scanning could be divided into three
sections:
a.

Total scanning–how much scanning did the subjects
devote to the diVerent faces in general? For each subject we calculated the total number of Wxations and
total scan time (the sum of the durations of all Wxations) used to scan each face. We also analyzed the
mean duration of individual Wxations to determine if
any change in total duration was due to an increase in
number of Wxations or the length of Wxations.

203

b. Regional distribution of scanning—where did they
scan within the faces? We divided each face into eight
diVerent feature regions: brow, left eye, right eye, nose,
mouth, chin, right cheek, left cheek (Fig. 2) (Groner
et al. 1984). Approximate criteria were used uniformly
to draw the boundaries between regions for diVerent
faces. Fixations were classiWed according to their
region. As a secondary comparison we contrasted
upper (brow and eyes) with lower face (nose, mouth,
chin, and cheeks), and right hemi-face (right eye and
cheek) with left hemi-face (left eye and cheek). As with
total scanning, we analyzed both number of Wxations
and scan time per region.
c. Scanning dynamics—how did they scan faces? To analyze the sequence of Wxations between diVerent facial
regions, we performed two separate analyses. Both
were directed at determining the variability of scanning
sequences.
The Wrst analysis examined the local predictability of
scanning using Markov Wrst-order matrices (Rizzo
et al. 1987; Gbadamosi and Zangemeister 2001). We
created a variable called ‘pairs’, where the Wrst digit
represented the facial region where the prior Wxation
had been located, and the second digit the facial region
of the current Wxation. With eight diVerent regions, this
creates 64 possible pairs. These can be formatted as an
8 £ 8 Wrst-order Markov matrix, in which the rows
represent the Wrst Wxation and the columns represent
the second Wxation. The asymmetric lambda (aL) is a
measure of the predictability or non-random nature of
the pairs within this matrix.
One particular sequence pair of interest is the ‘locally
repetitive pair’. These are pairs where both Wxations
are in the same region. Whether these locally repetitive
Wxation pairs should be included in analyses of scan
sequencing is contentious. Others have used ‘compressed’ Markov matrices that omit locally repetitive
pairs (Gbadamosi and Zangemeister 2001). We too
used compressed Markov matrices, which better assess
the inter-regional Xow of information processing. This
yielded compressed asymmetric lambda values (aLc).
We calculated eight aLc values for each subject, grouping data with common orientation (upright versus
inverted) and fame (novel versus famous). Our prior
report on normal subjects found no signiWcant eVect of
morphing on aLc values (Barton et al. 2006), so we
eliminated this as a main eVect from the analysis.
We also performed a separate analysis of locally repetitive pairs. These may represent information gathering
on a diVerent scale, possibly directed at local component analysis rather than perception of the whole structure. We computed the frequency of such pairs, in the

123

204

Exp Brain Res (2007) 181:199–211

Fig. 2 Top shows an example
of facial images used, in this
case a morph series between
Demi Moore and Julia Roberts.
Bottom shows an example of facial regions. Each face is segmented into eight regions: brow,
right eye (of the depicted face,
not the viewer’s hemiWeld), left
eye, nose, mouth, chin, right
cheek, and left cheek

eight diVerent conditions varying in orientation, morphing, and fame.
The second analysis examined the global predictability of
scanning. That is, how much did the whole scanning
sequence for one face resemble that of another by the
same subject? Recent scanning studies (Brandt and Stark
1997; Gbadamosi and Zangemeister 2001) have borrowed
techniques from DNA sequence analysis to attempt to
characterize this. The strategy is to quantify how much
change is required to transform one sequence into another.
One may make insertions, deletions, or substitutions, with
the latter counting twice, since a substitution is the equivalent of a combined insertion and deletion. Here too we
used a compressed variant that eliminated consecutive repetitions of a single area (Gbadamosi and Zangemeister
2001). The number of changes is divided by the average
number of items (Wxations) in the two sequences being
compared, to give the number of changes per Wxation.
This is called the ‘editing cost’. With this formula, the
minimum editing cost of ‘0’ indicates identical sequences,
whereas the maximum editing cost of ‘2’ equals completely dissimilar sequences.
d. Statistical analysis
We used ANOVA with repeated measures to analyze
the eVects of the diVerent factors, using JMP IN 5.1.2
(SAS Institute, www.jmpin.com). Where suYcient
data was available (e.g., total scanning), this was performed within the two subjects, using orientation,
fame, and morphing as the main eVects, across the
eight diVerent series of faces. Otherwise, for all parameters, the data of each subject was contrasted against
the controls independently, with subject group (prosopagnosic patient versus the controls) as an additional
main eVect. The normal data sometimes showed negligible eVects of certain factors on certain variables (e.g.,
morphing and fame for aLc values in the Markov
matrix analysis). Where this was present, those factors
were eliminated as main eVects in the ANOVA analysis.
SpeciWc planned comparisons for upright faces and

123

upright unmorphed faces were examined using linear
contrasts within the ANOVA analysis, and are reported
as P-values derived from t-tests.

Results
Total scanning
As we previously reported, normal subjects use fewer Wxations and less viewing time to make identity judgments
with upright faces than they do with inverted faces (Fig. 3).
They also require fewer Wxations and less viewing time to
make such judgments about famous faces than they do with
novel faces, and this eVect for fame is similar for both
upright and inverted faces.
Subject 005(A) did not show any eVect of orientation.
He did show a trend to a main eVect of fame on the number
of Wxations (9.01 Wxations/face for famous versus
10.21 Wxations/face for novel, F(1,112) = 3.21, P < 0.077),
but not on scan time. Planned contrasts between famous
and novel faces showed similar trends for number of Wxations for all unmorphed faces (7.75 Wxations/face for
famous versus 10.56 Wxations/face for novel, t = 1.77,
P < 0.079) and for upright unmorphed faces (7.5 Wxations/
face for famous versus 11.25 Wxations/face for novel,
t = 1.67, P < 0.098). Compared to the normal subjects, scan
times for faces in general were signiWcantly longer for subject 005(A) (4,048 ms/face) than for the controls (1,934 ms/
face, F(1,7)=22.8, P < 0.002).
Analysis of the duration of individual Wxations (rather
than total scan time) showed that these were signiWcantly
longer for subject 005(A) (415 ms/Wxation) than for
controls (242 ms/Wxation, F(1,940) = 71.1, P < 0.0001).
Also, the interaction between fame and orientation diVered
between subject 005(A) and controls [F(1,940) = 6.94,
P < 0.009]: subject 005(A) had longer individual Wxations
with upright famous faces than with inverted famous faces
or with novel faces, whereas controls showed no such relative increase for upright famous faces.

Exp Brain Res (2007) 181:199–211

205

Fig. 3 Total scanning. Left
graphs are for number of Wxations and right graphs show scan
time, summed from the durations of all Wxations used to scan
each face. Control data is shown
in the top graphs, with error
bars representing one standard
error. Solid bars are for unmorphed faces, striped bars for morphed faces. Subject 005(A) is
shown in the middle graphs and
subject 002(D) in the bottom
graphs. Both subjects show tendencies to less scanning for famous faces, though this was not
signiWcant for subject 005(A).
Subject 005(A) does not show
the normal orientation eVect of
less scanning with upright faces,
while subject 002(D) shows an
anomalous eVect of more scanning of upright faces

Subject 002(D) showed a slight eVect of orientation on
Wxation number [F(1,112) = 4.29, P < 0.041]—paradoxically with more Wxations on upright (8.61 Wxations/face)
than inverted faces (6.77 Wxations/face)—but not on scan
time. There was a signiWcant eVect of fame on scan time
(1,949 ms/face for famous versus 2,743 ms/face for novel,
F(1,112) = 14.2, P < 0.0003) but not on number of Wxations. A planned contrast showed signiWcantly shorter scan
time for famous faces (1,874 ms/face) than novel faces
(3,054 ms/face) in the upright unmorphed condition
(t = 2.10, P < 0.038). The contrast with controls showed
that subject 002(D) was not signiWcantly diVerent from
controls in either number of Wxations or scan time per face.
Subject 002(D) did show a signiWcant diVerence in the
eVect of orientation on both Wxation number [F(1,940) =
11.52, P < 0.0007] and scan time [F(1,940) = 4.57,
P < 0.033]: while controls scanned upright faces less than
inverted faces, she showed the opposite pattern. Analysis of
the duration of individual Wxations showed, as with subject
005(A), that these were signiWcantly longer for subject
002(D) (392 ms/Wxation) than for controls [242 ms/Wxation, F(1,940) = 47.19, P < 0.0001]. There was also a signiWcant diVerence in the eVect of fame [F(1,940) = 18.8,
P < 0.0001]: while controls showed equivalent durations of
Wxations for famous faces and novel faces (243 ms versus
235 ms, t = 0.81, P = n.s.), subject 002(D) had longer Wxation

durations on novel faces than on famous faces (434 ms versus 328 ms, t = 4.3, P < 0.0001).
To summarize, subject 005(A) and subject 002(D)
required longer scan times to process faces, in large part not
because they used more Wxations, but because their Wxations
lasted longer. The presence of an orientation-dependent
expert face mechanism should be revealed as orientation
eVects. Subject 005(A) showed no orientation eVects and
subject 002(D) showed an anomalous one, with more Wxations on upright than inverted faces. The presence of residual face recognition should be revealed as fame eVects.
Subject 005(A) showed a trend to a normal fame eVect on
Wxation number but not scan time, and subject 002(D) had a
normal fame eVect on scan time but not Wxation number.
Both also showed anomalous fame eVects on the duration
of individual Wxations, with subject 005(A) showing longer
Wxations on upright famous faces and subject 002(D)
showing longer Wxations on novel faces.
Regional distribution of scanning
Our previous analysis of the control data (Barton et al.
2006) showed that the interaction of facial region and fame
was modest, but the interaction between facial region and
orientation was highly signiWcant (Fig. 4). Normal subjects
scanned the upper face more than the lower in upright

123

206

Exp Brain Res (2007) 181:199–211

Fig. 4 Regional distribution of
scanning. Conventions as in
Fig. 3. Upright face data are
plotted alongside inverted face
data. Normal subjects scan the
eyes, nose, and mouth the most,
and show a bias for the eye in the
left hemiWeld (right eye in
upright faces and left eye in
inverted faces). Subject 005(A)
shows a similar distribution, but
without the left hemiWeld bias.
Subject 002(D) has an
anomalous distribution,
focusing more on the brow and
cheek, as well as a right
hemiWeld bias

Fig. 5 Lateralized distribution of scanning. Left graphs are control
subjects, with one standard error shown in error bars, middle graphs
are for subject 005(A), and right graphs are for subject 002(D). Top
graphs compare Wxation in the left versus the right hemiWeld, summing
scanning from eye and cheek regions. Subject 005(A) has lost the
normal left hemiWeld bias, and subject 002(D) has an anomalous right

hemiWeld bias. Bottom graphs compare Wxation in the upper (eyes and
brow) versus the lower (nose, mouth, chin, and cheeks) face. Normal
subjects Wxate the upper face more in upright faces and the lower face
more in inverted faces. Subject 005(A) shows a similar pattern, but
subject 002(D) Wxates the lower face more regardless of orientation

faces, but the lower face more than the upper when the face
was inverted (Fig. 5B). Furthermore, they tended to scan
the side of the face in the left hemiWeld more than the side
in the right hemiWeld, regardless of orientation (Fig. 5A).

The comparison of subject 005(A) with the controls
showed, for number of Wxations, a trend to a signiWcant
overall diVerence in face parts scanned [F(1,441) = 1.96,
P < 0.060], and a signiWcant interaction between group,

123

Exp Brain Res (2007) 181:199–211

orientation and face part [F(1,441) = 3.51, P < 0.0012). For
scan time subject 005(A) diVered signiWcantly from
the controls in face parts scanned [F(1,441) = 7.09,
P < 0.0001], and a signiWcant interaction between group,
orientation, and face part [F(1,441) = 4.67, P < 0.0001].
The general distribution of subject 005(A)’s scanning over
facial regions was not that dissimilar to controls, though,
with the main diVerence being that he did not show the normal bias for the left hemiWeld (Fig. 5). His division of Wxations between the upper versus lower face was very similar to
the controls, with more upper than lower Wxations in upright
faces and more lower than upper Wxations in inverted faces.
The comparison of subject 002(D) with the controls
showed a signiWcant diVerence in face parts scanned for
both Wxation number [F(1,441) = 16.48, P < 0.0001] and
scan time [F(1,441) = 13.15, P < 0.0001]. It also conWrmed
a signiWcant interaction between group, face region and orientation [F(1,441) = 10.84, P < 0.0001]. Subject 002(D)
had highly anomalous scan distribution, with an emphasis
on regions that are not highly salient for controls, such as
the left cheek and brow (Fig. 4). Her scanning was also
highly skewed toward the right hemiWeld, and she did not
show the normal concentration of the upper face and eye
region in upright faces (Fig. 5).
Scanning dynamics
Markov and local repetitions
In the controls, the compressed lambda showed an orientation eVect, with local scanning transitions being more random in inverted faces than in upright faces, but no eVect of
fame. Comparisons of subject 005(A) and subject 002(D)
with the controls showed no signiWcant diVerences. Subject
005(A) did show the normal inversion eVect for aLc, and
subject 002(D) had a small inversion eVect for novel faces
at least.
For local repetitions, the controls show no signiWcant
eVects of orientation or fame (Table 1). The comparisons of
subject 005(A) and subject 002(D) with the controls
showed that subject 005(A) had similar frequencies of local
repetitions, while subject 002(D) had almost twice as many
local repetitions as the controls [F(1,49) = 5.92, P < 0.046].
String editing costs
Normal subjects show eVects of both fame and orientation,
as well as a signiWcant interaction between the two. Editing
costs are least for novel upright faces, suggesting that scanning is most stereotyped for these faces (Fig. 6).
Subject 005(A) also showed a signiWcant eVect of fame
[F(1,112) = 18.04, P < 0.0004], but no signiWcant eVect of
orientation, or interaction with orientation. The eVect of

207

fame in subject 005(A) was anomalous, though, with less
editing costs for famous than for novel faces. The comparison with normal subjects conWrmed a signiWcant diVerence
in the eVect of fame [F(1,138) = 22.8, P < 0.0001]: while
the controls had higher editing costs with famous than novel
faces (t = 5.02, P < 0.0001), subject 005(A) had higher editing costs with novel faces than famous faces (t = 3.44,
P < 0.0007). Linear contrasts between subject 005(A) and
the controls showed that he had only slightly diVerent editing costs for famous faces (t = 2.02, P < 0.045) but much
higher editing costs for novel faces (t = 5.04, P < 0.0001).
Subject 002(D) showed no signiWcant eVects of fame or
orientation. The comparison with the controls showed that
subject 002(D) had much higher editing costs than the controls for all stimuli, regardless of fame or orientation
[F(1,138) = 16.91, P < 0.007].
To summarize the temporal dynamics, the local structure
of scanning (as indexed by the Markov and local repetition
analyses) showed an orientation eVect in subject 005(A) and
possibly a blunted one in subject 002(D); furthermore, subject
002(D) had an abnormally large number of local repetitions.
Global scan structure, as indexed by the string editing analysis, was highly chaotic in subject 002(D) and not modulated
by familiarity or orientation. The expertise eVect in subject
005(A) is also muted, in that he failed to show a signiWcant
eVect of orientation. However, he had an anomalous familiarity eVect, with less stereotyped scanning of novel faces than
famous faces, the reverse of the situation in the controls. This
was due to the fact that his global scan patterns for novel faces
were more random than those of controls, whereas his editing
costs for famous faces were similar to those of controls.

Discussion
The results show signiWcant diVerences in face scan patterns
between the prosopagnosics and the controls. Furthermore,
Table 1 Local scanning transitions
Control
Mean

SD

DM

KN

0.37

0.12

0.38

0.27

Asymmetric lambda
Famous

Upright

Famous

Inverted

0.31

0.12

0.24

0.30

Novel

Upright

0.35

0.11

0.37

0.32

Novel

Iinverted

0.26

0.07

0.21

0.28

0.25

0.10

0.34

0.40

Local repetitions
Famous

Upright

Famous

Inverted

0.23

0.07

0.35

0.45

Novel

Upright

0.26

0.08

0.34

0.39

Novel

Iinverted

0.23

0.06

0.30

0.46

123

208

Exp Brain Res (2007) 181:199–211

Fig. 6 Global scan patterns.
String editing costs are shown
against each of the morphed
images in a series, ranging from
9 to 21, where 21 is 100% of the
other face in a pair, and 11 is the
50:50 mixture. Control data at
the top left show that editing
costs are least for novel upright
faces, suggesting a more
stereotyped scanning of these
faces. Error bars show one
standard error for two of the
series. At bottom left, subject
005(A) shows editing costs that
are similar to controls for the
famous upright faces, but higher
editing costs than the controls
for the novel upright faces. He
also has a reduced inversion
eVect for the novel faces.
Subject 002(D) shows very high
editing costs for all faces,
without consistent eVects of
orientation or fame

they also show diVerences between the two prosopagnosics,
suggesting that scan patterns might reXect not only eVects
attributable to the absence of face recognition but also
eVects related to the type of prosopagnosia.
Besides the eVects of orientation and fame, which were
the main foci of this report, a few general eVects emerged
from the data. Face viewing is more diYcult for subjects
002(D) and 005(A) in general, as shown by their longer
scan time, which was due in large part to longer Wxation
durations. Hemianopia may have accentuated this diYculty
in subject 005(A), but subject 002(D) has no such Weld
defect. Normal subjects also showed a concentration of
scanning in the left hemiWeld, regardless of orientation.
However, subject 005(A) showed no hemiWeld asymmetry
in scanning. This is not likely due to his hemianopia, as a
left hemianopia should if anything lead to an adaptive
increase in the scanning of the left hemiWeld, rather than a
reduction (Behrmann et al. 1997; Barton et al. 1998). Subject 002(D), who does not have a hemianopia, has a pronounced bias to the opposite right hemiWeld. Previous
studies of normal subjects have also shown left hemiWeld
biases in at least some scanning parameters (Gallois et al.
1989; Mertens et al. 1993; Phillips and David 1997).
Whether this hemiWeld asymmetry reXects a predominant
role for the right hemisphere in processing faces has been
argued (Grega et al. 1988), but the loss of the left bias in

123

subject 005(A), a patient with a visible right-sided lesion, is
of interest in this debate.
Evidence for a residual expert mechanism
Is there evidence that subject 005(A), after a lesion of the
right fusiform face area, has residual facial expertise? He
does not show an orientation eVect on total scanning time
or number of Wxations. While the controls scanned
upright faces less, presumably because they are more
eYcient at processing upright faces, subject 005(A)
showed no diVerence between upright and inverted faces.
However, the distribution of scanning by subject 005(A)
is relatively normal and does show the normal pattern of
a bias for the upper face in upright faces and a bias for the
lower face in inverted faces. In scanning dynamics, transitions from one Wxation to another were less random in
upright than in inverted faces, for both controls and subject 005(A). However, his global scan patterns showed
only a trend to be more stereotyped for novel faces in the
upright than inverted presentation, with no signiWcant
orientation eVect. Furthermore, his global scan patterns
with novel upright faces are more chaotic than those of
the controls. Others have argued that the stereotyped
scanning of novel faces reXects the guidance of a generic
face schema possessed by the expert mechanism (Rizzo

Exp Brain Res (2007) 181:199–211

et al. 1987). If so, subject 005(A)’s more chaotic scanning of novel faces may reXect degraded expert processing, consistent with his demonstrated diYculties in
processing facial conWgurations. Along with the loss of
the normal left hemiWeld bias, all of these Wndings
suggest that subject 005(A)’s lesion has degraded but not
completely eradicated all traces of his expert face processing mechanism.
Subject 002(D), with developmental prosopagnosia, has
more profound and anomalous changes. In terms of orientation eVects, she scans upright faces more than inverted
ones, the opposite of normal. Her scanning of upright and
inverted faces both concentrate on the lower face, whereas
controls switch from emphasizing the upper face in upright
stimuli to the lower face in inverted stimuli. She shows no
eVect of orientation in her global scan patterns, which in
controls are more chaotic for inverted faces. In other
eVects, she has a highly anomalous distribution of scanning, spending more time on less salient regions such as
the cheek and brow. She makes far more local repetitions
of Wxations. Her global scan patterns are all highly chaotic.
The overall impression is that subject 002(D) shows no
evidence of normal face-expert mechanisms in her scanning of faces: she lacks normal orientation eVects, scans
anomalous regions of the face, and has highly chaotic scan
patterns.
Thus subject 005(A) with an acquired unilateral lesion
appears to have some residual face-expertise mechanisms,
while subject 002(D) with developmental prosopagnosia
has virtually none. The relatively preserved distribution of
subject 005(A)’s Wxations is consistent with conclusions of
normal Wxation distribution in another patient with acquired
prosopagnosia (Le et al. 2003). In contrast, subject 002(D)
has particularly chaotic scanning, with disorganized global
sequences and anomalous distributions of Wxations that do
not mirror the normal distribution of regional saliency
within faces (Shepherd et al. 1981). This too, is similar to a
previous study of four patients with hereditary prosopagnosia, which showed a tendency of these subjects to Wxate
external facial elements rather than the internal features
(Schwarzer et al. 2006). Subject 002(D)’s de-emphasis of
the eye region and increased scanning of less salient brow
and cheek regions is also reminiscent of other reports of
anomalous scanning in other patient groups. These include
the Wnding that autistic subjects Wxate the mouth more than
the eyes (Klin et al. 2002), and that schizophrenic subjects
Wxate salient facial features less (Loughland et al. 2002).
The data of subject 002(D) raise the possibility that these
other Wndings might reXect an innate disorganization of
face processing. However, caution is required for such
claims, as there are other possible explanations for such
anomalous scan patterns, including avoidance of social
engagement, for example.

209

Evidence for residual covert familiarity
Fame eVects would indicate the existence of residual facial
memories and their ability to inXuence data acquisition
during face viewing. This residual ability may be considered a type of covert face recognition in that, while prosopagnosic patients deny recognizing faces, inXuences of facial
familiarity can still be found in their behavior or perceptual
decisions (Bruyer 1991; Young 1994).
Subject 005(A) showed a trend to the normal eVect of
fewer Wxations on famous faces, but tended to take more
time with each Wxation on upright famous faces, an anomalous eVect. His global scan patterns also showed a signiWcant diVerent between famous and novel faces, but this too
was anomalous, in that novel faces were more chaotically
scanned than famous faces. However, this was due to novel
faces being scanned more randomly by subject 005(A) than
by controls, whereas his global scan patterns for famous
faces were comparable to controls. We suggest that these
results might indicate that subject 005(A)’s scanning of
famous faces are still guided by residual facial memories.
Previous studies have shown that he has reasonably good
imagery for famous faces, suggesting that his facial memories may be relatively spared (Barton and Cherkasova
2003). This contrasts with his more chaotic scanning of
novel faces, which may reXect impaired perceptual processing and access to a generic facial schema.
Subject 002(D) showed fewer familiarity eVects than
subject 005(A). She did spend less overall scan time viewing famous faces, as did normal controls. She also made
shorter Wxations on famous faces, whereas normal subjects
did not show an eVect of fame on the duration of Wxations.
Of note, subject 002(D) showed no signiWcant eVect of
fame in her global scan patterns.
Thus the eVects of fame are more prominent in subject
005(A) than in subject 002(D). Subject 005(A) has global
scan patterns for famous faces that are similar in their
degree of stereotypy to those of controls, and less chaotic
than his scanning of novel faces. This suggests that the
presence of facial memories in subject 005(A) can organize
his scanning of famous faces in a manner similar to controls, even though his scanning of novel faces is more disorganized. However, subject 005(A) also diVers from
controls in that his Wxations on famous faces last longer.
Longer Wxations can be found when perceptual information
is degraded artiWcially (van Diepen 1995). In subject
005(A), it may be that weak matches between his pathologically degraded percepts and intact facial memories generate a drive to increase data acquisition time during each
Wxation, more than occurs for novel faces where no strong
memories exist.
These Wndings provide interesting parallels to our other
behavioral data on covert recognition. Patients with unilateral

123

210

right occipitotemporal lesions had the strongest evidence of
covert recognition with either direct forced-choice or indirect priming techniques (Barton et al. 2001a, 2004). These
covert eVects correlated with their residual imagery for faces,
which was also best preserved in patients with right occipitotemporal lesions (Barton and Cherkasova 2003). In contrast,
patients with childhood-onset prosopagnosia had little
covert recognition or facial imagery (but see (Michelon
and Biederman 2003) for a report of a developmental prosopagnosic with residual face imagery). We postulated
Wrst, that residual facial memories may be a prerequisite for
covert recognition, and second, that these may be lacking in
childhood-onset prosopagnosia, as any failure in the complex hierarchy of processes in face recognition would preclude the establishment of facial memories after early-onset
lesions. In this report we Wnd stronger eVects of fame on
scan patterns in a subject with unilateral right occipitotemporal lesion, with good evidence of covert ability and
facial imagery in our prior reports, and weaker fame eVects
(chieXy a diVerence in total scan time) in the patient with
developmental prosopagnosia.
In summary, investigations of the scan patterns in these
two patients reveal degradation of expert face processing
in both patients, in whom prior testing had shown severe
impairments in the perceptual processing of facial conWguration. This degradation was more profound in subject
002(D), who had a developmental prosopagnosia. However, residual eVects of familiarity were still evident in the
global scan patterns of subject 005(A), who has evidence
of good facial imagery and covert recognition, but not in
subject 002(D). These Wndings show how the distribution
and organization of scanning Wxations can yield data about
the information processing of faces in prosopagnosic
patients.
Acknowledgments JB was supported by a Canada Research Chair,
a Michael Smith Foundation for Health Research Senior Scholarship,
CIHR MOP-77615 and NIMH R01-MH069898. James Intriligator
assisted with experimental programming and Rebecca Hefter with data
analysis.

References
AlthoV R, Cohen N (1999) Eye-movement-based memory eVect: a
reprocessing eVect in face perception. J Exp Psychol Learn Mem
Cogn 25:997–1010
Bartlett J, Searcy J (1993) Inversion and conWguration of faces. Cognit
Psychol 25:281–316
Barton J, Cherkasova M (2003) Face imagery and its relation to perception and covert recognition in prosopagnosia. Neurology
61:220–225
Barton J, Behrmann M, Black S (1998) Ocular search during line
bisection. The eVects of hemi-neglect and hemianopia. Brain
121:1117–1131

123

Exp Brain Res (2007) 181:199–211
Barton J, Cherkasova M, O’Connor M (2001a) Covert recognition in
acquired and developmental prosopagnosia. Neurology 57:1161–
1167
Barton J, Keenan J, Bass T (2001b) Discrimination of spatial relations
and features in faces: eVects of inversion and viewing duration.
Brit J Psychol 92:527–549
Barton J, Press D, Keenan J, O’Connor M (2002) Lesions of the fusiform face area impair perception of facial conWguration in prosopagnosia. Neurology 58:71–78
Barton J, Cherkasova M, Press D, Intriligator J, O’Connor M (2003a)
Developmental prosopagnosia: a study of three patients. Brain
Cogn 51:12–30
Barton J, Deepak S, Malik N (2003b) Attending to faces: change detection, familiarisation and inversion eVects. Perception 32:15–28
Barton J, Cherkasova M, Hefter R (2004) The covert priming eVect of
faces in prosopagnosia. Neurology 63:2062–2068
Barton J, RadcliVe N, Cherkasova M, Edelman J, Intriligator J (2006)
Information processing during face recognition: the eVects of
familiarity, inversion and morphing on scanning Wxations. Perception 353:1089–1105
Behrmann M, Watt S, Black S, Barton J (1997) Impaired visual search
in patients with unilateral neglect: an oculographic analysis. Neuropsychologia 35:1445–1458
Brandt S, Stark L (1997) Spontaneous eye movements during visual
imagery reXect the content of the visual scene. J Cogn Neurosci
9:27–38
Bruyer R (1991) Covert facial recognition in prosopagnosia: a review.
Brain Cogn 15:223–235
Deco G, Schurmann B (2000) A neuro-cognitive visual system for object recognition based on testing of interactive attentional topdown hypotheses. Perception 29:1249–1264
Farah M, Wilson K, Drain M, Tanaka J (1998) What is “special” about
face perception? Psychol Rev 105:482–498
Gallois P, Buquet C, Charlier J, Paris V, Hache J-C, Dereux J-F (1989)
Asymmetry in the strategy for visual perception of faces and emotional facial expression. Rev Neurol 145:661–664
Gbadamosi J, Zangemeister WH (2001) Visual imagery in hemianopic
patients. J Cogn Neurosci 13:855–866
Grega D, Sackheim H, Sanchez E, Cohen B, Hough S (1988) Perceiver
bias in the processing of human faces: neuropsychological mechanisms. Cortex 24:91–117
Groner R, Walder F, Groner M (1984) Looking at faces: local and global aspects of scanpaths. In: Gale A, Johnson F (eds) Theoretical
and applied aspects of eye movement research. Elsevier, Amsterdam, pp 523–533
Henderson J (2003) Human gaze control during real-world scene perception. Trends Cogn Sci 7:498–504
Kemp R, McManus C, Pigott T (1990) Sensitivity to the displacement
of facial features in negative and inverted images. Perception
19:531–543
Klin A, Jones W, Schultz R, Volkmar F, Cohen D (2002) Visual Wxation patterns during viewing of naturalistic social situations as
predictors of social competence in individuals with autism. Arch
Gen Psychiatry 59:809–816
Le S, Raufaste E, Demonet JF (2003) Processing of normal, inverted,
and scrambled faces in a patient with prosopagnosia: behavioural
and eye tracking data. Brain Res Cogn Brain Res 17:26–35
Leder H, Bruce V (1998) Local and relational aspects of face distinctiveness. Q J Exp Psychol A 51:449–473
Loughland CM, Williams LM, Gordon E (2002) Visual scanpaths to
positive and negative facial emotions in an outpatient schizophrenia sample. Schizophr Res 55:159–170
Malcolm G, Leung C, Barton J (2004) Regional variation in the inversion eVect for faces: diVerent patterns for feature shape, spatial
relations, and external contour. Perception 33:1221–1231

Exp Brain Res (2007) 181:199–211
Mertens I, Siegmund H, Grusser OJ (1993) Gaze motor asymmetries
in the perception of faces during a memory task. Neuropsychologia 31:989–998
Michelon P, Biederman I (2003) Less impairment in face imagery than
face perception in early prosopagnosia. Neuropsychologia
41:421–441
Phillips M, David A (1997) Viewing strategies for simple and chimeric
faces: an investigations of perceptual bias in normals and schizophrenic patients using visual scan paths. Brain Cogn 35:225–238
Rizzo M, Hurtig R, Damasio A (1987) The role of scanpaths in facial
recognition and learning. Ann Neurol 22:41–45
Rybak I, Gusakova V, Golovan A, Podladchikova L, Shevtsova N
(1998) A model of attention-guided visual perception and recognition. Vision Res 38:2387–2400
Schwarzer G, Huber S, Gruter M, Gruter T, Gross C, Hipfel M,
Kennerknecht I (2006) Gaze behaviour in hereditary prosopagnosia. Psychol Res [epub]
Searcy J, Bartlett J (1996) Inversion and processing of component and
spatial-relational information in faces. J Exp Psychol 22:904–915

211
Sergent J (1984) An investigation into component and conWgurational
processes underlying face perception. Brit J Psychol 75:221–242
Shepherd J, Davies G, Ellis H (1981) Studies of cue saliency. In:
Davies G, Ellis H, Shepherd J (eds) Perceiving and remembering
faces. Academic, London, pp 105–131
Stacey P, Walker S, Underwood J (2005) Face processing and familiarity: evidence from eye-movement data. Brit J Psychol 96:407–422
van Diepen P (1995) Chronometry of foveal information extraction
during scene perception. In: Findlay J, Walker R, Kentridge R
(eds) Eye movement research: mechanisms, processes and applications. Elsevier, Amsterdam, pp 349–362
Yin R (1970) Face recognition: a dissociable ability? Neuropsychologia 23:395–402
Young A (1994) Covert recognition. In: Farah M, RatcliV G (eds) The
neuropsychology of high-level vision. Lawrence Erlbaum Associates, Hillsdale, pp 331–358
Zangemeister W, Sherman K, Stark L (1995) Evidence for a global
scanpath strategy in viewing abstract compared with realistic images. Neuropsychologia 33:1009–1025

123