The Impact of Image Based Factors and Training on
Threat Detection Performance in X-ray Screening
Adrian Schwaninger∗, Anton Bolfing∗, Tobias Halbherr†, Shaun Helman‡, Andrew Belyavin‡ and Lawrence Hay‡
∗Max Planck Institute for Biological Cybernetics, Tu¨bingen, Germany, and
Department of Psychology, University of Zurich, Switzerland
e-mail: a.bolfing@psychologie.uzh.ch, a.schwaninger@psychologie.uzh.ch
†Department of Psychology, University of Zurich, Switzerland, Email: t.halbherr@psychologie.uzh.ch
‡QinetiQ Limited, Hampshire, United Kingdom
Abstract—In this study, two experiments are reported which
investigated the relative importance of five different image based
factors and one human factor (training) in mediating threat
detection performance of human operators in airport security
x-ray screening. Experiment 1 was based on a random sample
of roughly 16’000 records of threat image projection (TIP) data.
TIP is a software function available on state-of-the-art x-ray
screening equipment that allows the projection of fictional threat
images (FTIs) into x-ray images of passenger bags during the
routine baggage screening operation. Analysis of main effects
showed that image based factors can substantially affect screener
detection performance in terms of the hit rate (identification of
FTIs). There were strong effects of FTI view difficulty (rotation
of FTIs) and superposition of FTIs by other objects in the x-ray
image of a passenger bag. The amount of opacity in the x-ray
image of a passenger bag had a small although significant effect
on detection performance. The two image based factors clutter
and bag size did not have a significant effect.
Experiment 2 was conducted using an offline-test in order to
provide controlled and more detailed data for analyzing the
image based factors from Experiment 1, as well as the human
factor of training. In particular the individual factors’ main
effects on detection performance, main effects of all factors taken
together and factor interactions were analyzed. In the test design
the following image-based factors were varied systematically:
Threat (FTI) category (guns, knives, improvised explosive devices,
other threats), view difficulty, superposition, bag complexity (a
combination of opacity and clutter) and bag size. Data were
collected from 200 screening officers at five sites across Europe.
For screener training all five sites use the same computer-
based training system. Consistent with the results obtained in
Experiment 1, there were large main effects of threat (FTI)
category, view difficulty, and superposition. Again consistent with
Experiment 1, effects of bag complexity (opacity and clutter) and
bag size were much smaller. In addition to Experiment 1, the
number of computer based training (CBT) hours was available
for each security officer participating in the study. Training
turned out to be a key driver to improving threat detection
performance in x-ray screening and seemed to mediate the effects
of some image based factors.
Possible implications regarding the enhancement of human-
machine interaction in x-ray screening are discussed.
I. INTRODUCTION
Screening passenger bags for threat items using state-of-
the art x-ray machines is an essential component of airport
security. Previous work (Schwaninger, 2003b, Schwaninger,
Hardmeier, & Hofer, 2005, and Schwaninger, Michel, &
Bolfing, 2007) has identified image based factors that affect
human performance in x-ray screening tasks: object view
difficulty, superposition by other objects and bag complexity
(opacity and clutter). Recently the question has been raised
whether bag size could be another image based factor that
affects detection of threat items when visually inspecting x-ray
images of passenger bags. In this study we determined effects
and interactions of image based factors and human factors
(amount of recurrent computer-based training). In addition,
with empirically based conclusions regarding the importance
of the bag size variable, by itself as well as in relation
with other performance relevant factors, this study provided
the scientific basis for a political decision making process
regarding the improvement of aviation security.
Two experiments are reported. Experiment 1 is based on threat
image projection (TIP) data. Experiment 2 is based on an
off-line computer based test, which allows investigating the
combined effects of image-based factors, effects of training
as well as factor interactions. The use of these two methods
to answer the same research question will ensure that the
overall approach is complementary. Both methods have their
own strengths and weaknesses: TIP data give high ecological
validity but low experimental control; off-line computer based
tests using controlled stimuli allow more experimental control,
but less ecological validity. If both methods provide the same
answer to the research question, this can be taken as stronger
evidence that the findings are genuine, and not simply an
artefact of the particular method used.
The two experiments both follow the paradigm using computer
algorithms to estimate image based factors that influence threat
detection performance in x-ray screening. This paradigm was
developed at University of Zurich and presented at ICRAT
2006 in Belgrade (Bolfing, Michel, & Schwaninger, 2006a)
and published before (Schwaninger, Michel, & Bolfing, 2007;
Bolfing, Michel, & Schwaninger, 2006b; Schwaninger, Michel,
& Bolfing, 2005). None of these papers used TIP data for
analysis, which ensures high ecological validity. Experiment
2 is based on a much larger data set than the previous studies
augmenting reliability. The inclusion of bag size and training
as additional factors is completely novel within this paradigm.
Since threat detection performance in aviation security x-ray
THIRD INTERNATIONAL CONFERENCE ON RESEARCH IN AIR TRANSPORTATION                 FAIRFAX, VA, JUNE 1-4 2008
ISBN: 978-0-615-20720-9317
screening depends on the x-ray images but also on the human
screeners-the final decision makers-human factors should not
be neglected in a comprehensive model whose goal is to
explain the x-ray threat detection process.
A. Image Based Factors
Schwaninger (2003b) and Schwaninger, Hardmeier, and
Hofer (2005) have identified three image based factors which
affect threat detection by x-ray screeners: view difficulty,
superposition, and bag complexity (see figure 1).
Fig. 1. Illustration of the three basic image based factors suggested by
Schwaninger (2003b) and Schwaninger, Hardmeier, and Hofer (2005)
The concepts of these image based factors have been math-
ematically modeled (Schwaninger, Michel, & Bolfing, 2007,
see Bolfing, & Schwaninger, 2007 for the latest version). View
difficulty is modeled as a statistically calculable value between
0 and 1 named FTI view difficulty. Superposition and bag
complexity are modeled as image processing measurements
with bag complexity being split up into clutter and opacity.
The introduction of the image based factor bag size in this
study necessitated normalization of earlier implementations
of clutter and opacity regarding bag size. Formulae and
short descriptions of the underlying concepts are specified in
Bolfing, & Schwaninger (2007).
II. THREAT IMAGE PROJECTION (TIP) χ2 ANALYSIS:
EXPERIMENT 1
A. Method
1) Threat Image Projection (TIP) Data: In order to ensure
high ecological validity, we decided to analyze data from
threat image projection (TIP). TIP is a software function of
state-of-the-art x-ray screening equipment used at security
checkpoints in airports, nuclear power plants, navigation
docks etc. In aviation security TIP distinguishes between
cabin baggage screening (CBS) and hold baggage screening
(HBS). In CBS, guns, knives, improvised explosive devices
(IEDs) and other threats are subject to identification and
confiscation. In HBS, the focus rests mainly on IEDs and
dangerous goods such as gasoline containers or diver lamps.
The current investigation is confined to CBS. In CBS TIP,
fictional threat items (FTIs) are occasionally projected into
x-ray images of passenger bags during the routine baggage
screening operation. A sufficiently large sample of TIP
events allows statistically reliable measurements of detection
performance of human operators (x-ray screeners) on-the-job
(Hofer & Schwaninger, 2005) and thus with high ecological
validity.
The data basis of this study consists of a random sample
of 16’329 TIP events that have been routinely recorded on-
the-job with approximately 700 professional x-ray screeners
throughout the first half of 2007 at a large European airport.
We decided to apply χ2 analyses to each image based factor
separately to measure its impact on detection performance in
terms of hit rate (i.e. correctly judging a bag as being NOT
OK).
2) χ2 Analysis: To compare the effects on detection per-
formance of the independent variables1 FTI view difficulty,
superposition, opacity, clutter and bag size, the following
procedures were applied to the TIP data described above. A
histogram was created for each independent variable (image
based factor). For each variable the upper and lower 2.5% of
the cases in the data were excluded to remove outlier data from
the analysis. Furthermore this made possible the definition of
five equidistant bins with at least 100 data points each (TIP
events).
Hit rates were calculated for each of the five equidistant bins
to run χ2 tests with the null hypothesis H0 that the hit rates
are equal across bins. Effect size analysis based on Cohen
(1988) was used to compare the effect sizes of the different
independent variables. For detailed information on χ2 statistics
see for example Coolican (2004).
B. Results
The results below are listed separately for each image based
factor introduced above (see Bolfing, & Schwaninger, 2007
for further information and formulae). Each of the following
subsections begins with a graphical illustration of the image
based factors’ effects on the threat detection performance
measure hit rate. The x-axes show the five equidistant bins
into which the whole data range was subdivided. Low values
are on the left, high values on the right. The y-axes show
the hit rates of the image based factors’ bins. For reasons
of confidentiality hit rates cannot be given explicitly, but
the hit rate scales are reasonably chosen and kept constant
throughout the whole document.
Following the graphical illustrations (figures 2-6), statistical
test values are given in tables I-V. χ2 statistics can be
interpreted as follows: the larger the χ2(df, N) value the
larger the effect. Additionally χ2 effect sizes w are given.
1The variables correspond to the continuously represented variables used
in the multiple regression analysis in Experiment 2 (see figure 8)
THIRD INTERNATIONAL CONFERENCE ON RESEARCH IN AIR TRANSPORTATION                 FAIRFAX, VA, JUNE 1-4 2008
ISBN: 978-0-615-20720-9318
Again, the larger the effect size, the larger the effect.
However, please be aware that χ2 and w values do not state
the direction of the effect.
To summarize the χ2 analysis results a bar plot graphic is
provided at the end of this section illustrating the χ2 effect
sizes of the five image based factors on the hit rate (see
figure 7). The image based factors are arranged such that
their effects decrease in size.
1) FTI View Difficulty: Figure 2 illustrates the large impact
of FTI view difficulty on human detection performance in
terms of hit rate. This is partly due to the fact that objects
are more difficult when depicted from an unusual viewpoint
(see figure 1). Other factors contributing to this large impact
are the threat category of the object and the training of human
operators (see Experiment 2).
Fig. 2. Illustration of the impact of FTI view difficulty on hit rate.
TABLE I
χ2 ANALYSIS RESULTS: FTI VIEW DIFFICULTY
χ2 value χ2(4, N = 13′541) = 198.04
Significance Highly significant: p < .001
χ2 effect size w = .12
2) Superposition: Figure 3 illustrates the large effect of
superposition on detection performance.
Fig. 3. Illustration of the impact of superposition on hit rate.
3) Opacity: Figure 4 shows the significant but relatively
small influence of opacity on detection performance in terms
of hit rate.
TABLE II
χ2 ANALYSIS RESULTS: SUPERPOSITION
χ2 value χ2(4, N = 13′713) = 72.98
Significance Highly significant: p < .001
χ2 effect size w = .07
Fig. 4. Illustration of the impact of opacity on hit rate.
TABLE III
χ2 ANALYSIS RESULTS: OPACITY
χ2 value χ2(4, N = 13′718) = 9.90
Significance Significant: p < .05
χ2 effect size w = .03
Here the question arises whether it is opacity as a perceptual
concept that does not have much influence on threat detection
performance, or whether the image measurement formula of
opacity is not properly modeled.
4) Clutter: Figure 5 illustrates the hit rates of the five clut-
ter bins. There is no significant effect of clutter on detection
performance. As with opacity, the question arises whether it
is the concept of clutter that does not influence hit rates in
TIP, or whether the computational model of clutter needs to
be improved.
Fig. 5. Illustration of the impact of clutter on hit rate.
5) Bag Size: Figure 6 shows the effect of bag size on hit
rate in TIP.
As with clutter, the effect of bag size on detection
performance does not reach statistical significance.
THIRD INTERNATIONAL CONFERENCE ON RESEARCH IN AIR TRANSPORTATION                 FAIRFAX, VA, JUNE 1-4 2008
ISBN: 978-0-615-20720-9319
TABLE IV
χ2 ANALYSIS RESULTS: CLUTTER
χ2 value χ2(4, N = 13′726) = 0.98
Significance Not significant: p = .913
χ2 effect size w = .01
Fig. 6. Illustration of the impact of bag size on hit rate.
6) Comparison of the χ2 Effect Sizes: In figure 7, the effect
sizes w are compared. The factor FTI view difficulty has the
highest effect size with w = .12, while clutter shows the
lowest effect size with w = .01. The factors opacity, bag size
and clutter show small effect sizes. The effects of clutter and
bag size did not reach statistical significance.
Fig. 7. Comparison of the effect sizes among the image based factor.
C. Discussion
The results obtained in Experiment 1 are consistent with
earlier findings. Schwaninger, Hardmeier, and Hofer (2005)
found that viewpoint, superposition and bag complexity af-
fect screener performance. Schwaninger, Michel, and Bolfing
(2007) replicated these results. Using similar image mea-
surements as in Experiment 1, they measured similar effects
for FTI view difficulty, superposition, opacity (negatively
correlated with transparency in Schwaninger et al., 2007)
and clutter. However, several caveats are necessary to qualify
the appropriateness of the results obtained in Experiment 1.
Firstly, an analysis of auto-archive bags indicated that, as
would be anticipated, it is likely that TIP aborts are selectively
TABLE V
χ2 ANALYSIS RESULTS: BAG SIZE
χ2 value χ2(4, N = 13′758) = 4.45
Significance Not significant: p = .348
χ2 effect size w = .02
eliminating certain bags (e.g. small bags rather than large bags)
from the TIP image set, and thus reducing their presence.
Secondly, it is not always clear how closely aligned TIP scores
are with the specific operational situations encountered when
threats are deliberately hidden in difficult bags. But most
importantly, in Experiment 1 only main effects were analyzed.
In order to gain a more complete picture it is important to
conduct a more controlled experiment in which main effects
in combination and their interactions can be measured reliably.
This was conducted in Experiment 2.
III. OFF-LINE COMPUTER BASED TEST:
EXPERIMENT 2
A. Method
1) Participants: 200 X-ray screeners from five European
sites with varying amounts of training in x-ray image
interpretation.
2) Stimuli: The stimuli were 1024 complete threat images
(CTIs) and 1024 complete non-threat images (CNTIs). CTIs
were created by projecting fictional threat items (FTIs) into
1024 X-ray images of bags. FTIs for the study were eight
visually similar pairs of each of four types of threat items:
guns, knives, improvised explosive devices (IEDs), and
’other’ threats. Images of cabin baggage were captured from
x-ray machines at a European airport using the auto-archive
function. The images were revised by three airport security
supervisors to remove inappropriate images (e.g. images
containing more than one bag, images containing incomplete
bags, bags containing prohibited items or liquids, etcetera).
This procedure resulted in 7606 bag images. Additional
review by the QinetiQ team resulted in a total of 6659 bag
images from which the 1024 bags needed for the study were
drawn. The final 1024 bags used for the study were chosen
through a process of projecting the relevant FTIs into the
bags such that the variables of interest would be orthogonal
in the stimulus set. Several full sets of 2048 images (the
1024 images containing the FTIs, and the same images
without FTIs) were created. The one with the most desirable
properties in terms of variable orthogonality was chosen for
use in the study.
3) Design: The study employed a 4 (FTI category: guns,
knives, IEDs, other) x2 (view difficulty: easy, difficult) x2
(superposition: low, high) x2 (bag complexity: low, high) x2
(bag size: small, large) x2 (image type: FTI, no FTI) within-
participants design. Since there were 16 FTIs in each category,
this design results in a total of 16x4x2x2x2x2x2 = 2048 images
which were to be presented to the screeners. The images
THIRD INTERNATIONAL CONFERENCE ON RESEARCH IN AIR TRANSPORTATION                 FAIRFAX, VA, JUNE 1-4 2008
ISBN: 978-0-615-20720-9320
were presented to the screeners in a random order in multiple
testing sessions of 20 minutes each. As dependent variable
the detection performance measure d′ (Green & Swets, 1966)
was used. This measure provides a more valid estimate of
detection performance than the hit rate alone because it takes
the hit rate and the false alarm rate into account (see Hofer &
Schwaninger, 2004 for different measures of x-ray detection
performance). Since the off-line test showed each bag once
with a threat and once without one, accurate measurements of
hit and false alarm rates could be obtained.
B. Results
Data were analyzed in two ways. Firstly, by treating the
variables FTI view difficulty, superposition, opacity, clutter,
and bag size as continuous, a linear regression was employed
to assess the main effects of each image based factor on threat
detection performance separately. A multiple linear regression
was used to examine the main effects together. Additionally,
we calculated a linear regression with hours of recurrent
computer based training prior to testing as predictor. In order
to examine main effects as well as interactions between the
variables, the discrete variables FTI category, view difficulty,
superposition, bag complexity and bag size were used in an
analysis of covariance (ANCOVA). Training hours served as
covariate in the ANCOVA. Figure 8 shows the way in which
the continuous and discrete variables are related to each other.
Due to a high inter-correlation and a test design that demands
independence of its variables, opacity and clutter were encoded
into the single discrete variable bag complexity. FTI category
and view difficulty were encoded into a single continuous
variable because it is not sensible to encode either variable
directly into a continuous variable. Instead we defined the
variable FTI view difficulty as the difficulty-as measured in
threat detection performance (d′)-screening officers had in
solving a specific threat item in a specific view (easy or
difficult) across all other conditions (i.e. superposition, bag
complexity and bag size).
Fig. 8. Illustration of relationship between discrete and continuous repre-
sentations of variables
1) Linear Regression and Multiple Linear Regression:
The regression analyses will help us understand the direct
relationship between image based factors and d′, as well as
training hours and d′. Figure 9 shows the relative effect sizes,
the absolute values of the correlations with the dependent
variable d′, for the individual variables. For superposition
and training hours a logarithmic transformation was applied.
This transformation was necessary in order to achieve a linear
relationship between superposition and detection performance
d′. With .70, .63 and .58, FTI view difficulty, training hours
and superposition all have very high effect sizes. Opacity has
a moderate to small effect size with .22, clutter and bag size
have very small effect sizes with .05 and .07, respectively.
Except for clutter, all correlations are statistically significant.
Fig. 9. Illustration of effect sizes R
Figures 10 and 11 show the results of the multiple linear
regression with all image based factors: FTI view difficulty,
superposition (logarithmically transformed), opacity, clutter
and bag size. It shows the overall effect size, again the absolute
value of the correlation R, of all the image based factors
taken together. With R = 0.77 the effect size is very high.
The effect size of the only human factor analyzed (hours of
recurrent computer based training), with R = 0.63, is also
large. We can see that in the multiple linear regression model
the factor bag size is the only one not reaching statistical
significance. Put another way: In the presence of the other
image based factors bag size did not lead to a statistically
significant change in detection performance in our experiment.
As shown in figure 12 adding bag size to the linear model only
leads to a minimal increase of its effect size from R = 0.772
to R = 0.773.
Fig. 10. Multiple linear regression overview
2) ANCOVA: A repeated measures analysis of covariance
(ANCOVA) was conducted to analyze the main effects of
image based factors, their interactions and their interactions
with training. As can be seen in the main effects summary
of figure 13 the repeated measures ANCOVA leads to only a
slightly different pattern with regards to effect sizes than the
linear regression analyses. These differences are due to the
fact that, in contrast to the linear regression models, in the
THIRD INTERNATIONAL CONFERENCE ON RESEARCH IN AIR TRANSPORTATION                 FAIRFAX, VA, JUNE 1-4 2008
ISBN: 978-0-615-20720-9321
Fig. 11. Multiple linear regression details
Fig. 12. Combined effect size of image based factors and effect size of
training
ANCOVA analysis effects of the covariate training hours are
isolated from the effects of image based factors. Furthermore,
in the ANCOVA inter-individual differences between screen-
ing officers (’screener variance’) are taken into account. Su-
perposition shows the largest effect size (η2), followed by FTI
category, bag complexity and view difficulty. The main effect
of bag size is clearly smaller than the main effect of any other
image based factor. Training hours has noteworthy interactions
with FTI category and view difficulty. These interactions make
sense, since we know from other studies that training can
lead to comparatively larger performance increases for items
that are comparatively difficult for novices (Koller, Hardmeier,
Michel, & Schwaninger, in press)-for example improvised ex-
plosive devices (threat item category) or difficult views (view
difficulty). There is also a small interaction of training with
bag size, indicating that well trained screening officers are less
affected by effects of bag size. Figure 14 gives an overview
of the 10 largest interactions in the ANCOVA. All in all over
30 interactions reached statistical significance. Since the effect
sizes of most interactions are very small we decided only to
report interactions η2 ≥ .07. The interaction of view difficulty
with threat category can at least partly be explained by the fact
that detection performance of improvised explosive devices-
unlike guns or knives-is largely independent of viewpoint. The
interaction of superposition with view difficulty indicates that
with difficult viewpoints superposition plays a larger role in
determining detection performance than with easy views. The
interaction of superposition with threat category indicates that
some threat item categories are more sensitive to superposition
than others. For example, from the regression analysis above
we know that superposition effects are higher with knives than
with guns.
Fig. 13. Illustration of ANCOVA main effects and interactions with the
covariate training hours
Fig. 14. Illustration of the the ten largest ANCOVA interactions
C. Discussion
With an overall correlation of .77 the linear modeling of
detection performance with image based factors has a very
high explanatory power. Superposition, although not always
with the largest effect size, has shown the most robust effects
on detection performance. Interestingly and in contrast to what
one might have expected based on the results of the regression
analyses, the variable bag complexity (a combination of opac-
ity and clutter) showed a large effect size in the ANCOVA.
Apart from this, the ANCOVA results reflect the regression
analysis results closely, both in main effects and interactions.
Threat category and view difficulty had considerable interac-
tions with the covariate training hours. This shows that training
is particularly effective in the case of difficult item categories
such as IEDs and for difficult viewpoints. Bag size, although
intuitively plausible as relevant factor, turned out to play only
a minor role in determining threat detection performance. The
same is true for clutter.
IV. GENERAL DISCUSSION
There were large main effects of view difficulty and of FTI
difficulty in all of the analyses, as expected. The same was
true for superposition and complexity (to a bigger extent for
opacity than for clutter). Clearly, these factors need to be taken
account of in any future work on performance-relevant image
based factors. When looking at the influence on detection
performance of all image based factors together, there is no
statistically significant effect of bag size. When using a more
sophisticated model of data analysis including main effects of
THIRD INTERNATIONAL CONFERENCE ON RESEARCH IN AIR TRANSPORTATION                 FAIRFAX, VA, JUNE 1-4 2008
ISBN: 978-0-615-20720-9322
FTI view difficulty, superposition, bag complexity, bag size
and the interactions of these variables, there is a small effect
of bag size. In Experiment 2 we were also able to examine the
effect of the number of CBT training hours on threat detection
performance. The key finding from the study is that the
effect size for this variable was large, and seemed to mediate
the effect of some image based factors on threat detection.
Clearly, training is a key driver to improving threat detection
performance in x-ray screening, and more work needs to be
done to establish exactly which image based factors screeners
need to be trained in to give the best improvements in threat
detection accuracy.
V. RECOMMENDATIONS FOR IMPROVING
HUMAN-MACHINE INTERACTION IN X-RAY SCREENING
A. FTI View Difficulty and Superposition
The factor FTI view difficulty refers to the fact that the
identification of threat objects, as objects in general, is highly
dependent on their viewpoint as well as on properties of
the very object itself. Current x-ray screening equipment
provides only one x-ray image per passenger bag. More
recent technology can provide multiple views of a bag.
Figure 15 illustrates how such new systems might be able
to reduce the detection problems due to view difficulty and
superposition. Objects that are superimposed by other objects
from one perspective may be clearly visible from another
one. Furthermore, training is an important tool in lessening
detrimental effects on detection performance of difficult
views. Our ANCOVA analysis has supported earlier findings
that training leads to particularly large improvements in
detection performance for difficult views (Koller, Hardmeier,
Michel, & Schwaninger, in press).
Fig. 15. Illustrative example of how multi-view systems can help improving
detection performance in spite of undesirable view difficulty and superposition
effects.
B. Opacity
The image based factor Opacity refers to the amount of
opaque areas in an x-ray image. X-ray systems with higher
penetration have the potential to reduce detection problems
due to opacity. In addition, it is possible to implement image
measurement algorithms in x-ray equipment that warn the
human operator (x-ray screener) with a ”dark alarm”, which
would be triggered by opaque areas that are deemed too large
or dense for unassisted human interpretation. Manual search
would follow when a dark alarm was indicated.
C. Screener Selection and Training
A very important approach to face the problem of improving
threat detection performance in x-ray screening consists in
screener selection and screener training. The psychological
literature provides evidence that figure ground segregation
(related to superposition) as well as mental rotation (related to
view difficulty) are visual abilities that are fairly stable within
a person. For example Hofer, Hardmeier, & Schwaninger
(2006) and Hardmeier, Hofer, and Schwaninger (2006a) have
shown that using computer based object recognition tests in
a pre-employment assessment procedure can help to increase
detection performance of screeners substantially.
In addition to stable abilities, there are several aspects of visual
knowledge relevant to x-ray image interpretation. Knowledge
based factors such as knowing which objects are dangerous or
prohibited and what they look like in x-ray images are train-
able. Training also has beneficial effects on screeners’ abilities
to deal with certain image based factors. For example, training
particularly improves the ability to deal with difficult views.
Computer-based training can be a powerful tool to improve x-
ray image interpretation competency of screeners (e.g. Koller,
Michel, Hardmeier, & Schwaninger, in press; Schwaninger,
Hofer & Wetter, 2007; Ghylin, Drury, & Schwaninger, 2006).
ACKNOWLEDGMENT
This research was supported by the UK Department for
Transport, the European Civil Aviation Conference Technical
Task Force (ECAC TTF), and by the European Commission
Leonardo da Vinci Programme (VIA Project, DE/06/C/F/TH-
80403). Thanks to Zurich State Police, Airport Division and
all other security organizations, companies and airports that
supported this study by supplying screeners and data. Special
thanks to Olive Emil Wetter and Jonas Sourlier of VICOREG
for their valuable contributions in test construction, data anal-
ysis, and reporting.
REFERENCES
[1] S. Koller, D. Hardmeier, F. Hofer, and A. Schwaninger, “Investigating
training, transfer, and viewpoint effects resulting from recurrent cbt of
x-ray image interpretation,” in Journal of Transportation Security, in
press.
[2] A. Bolfing and A. Schwaninger, “Measurement formulae for image-
based factors in x-ray imagery,” in VICOREG Technical Report, Novem-
ber 26, 2007.
[3] A. Schwaninger, S. Michel, and A. Bolfing, “A statistical approach for
image difficulty estimation in x-ray screening using image measure-
ments,” in Proceedings of the 4th Symposium on Applied Perception
in Graphics and Visualization, ACM Press, New York, USA, 2007, pp.
123–130.
[4] D. Hardmeier, F. Hofer, and A. Schwaninger, “The role of recurrent cbt
for increasing aviation security screeners’ visual knowledge and abilities
needed in x-ray screening,” in Proceedings of the 4th International Avia-
tion Security Technology Symposium, Washington, D.C., USA, November
27 - December 1, 2006b, pp. 338–342.
THIRD INTERNATIONAL CONFERENCE ON RESEARCH IN AIR TRANSPORTATION                 FAIRFAX, VA, JUNE 1-4 2008
ISBN: 978-0-615-20720-9323
[5] ——, “Increased detection performance in airport security screening
using the x-ray ort as pre-employment assessment tool,” in Proceedings
of the 2nd International Conference on Research in Air Transportation,
ICRAT 2006, Belgrade, Serbia and Montenegro, June 24-28, 2006a, pp.
393–397.
[6] F. Hofer, D. Hardmeier, and A. Schwaninger, “Increasing airport security
using the x-ray ort as effective pre-employment assessment tool,”
in Proceedings of the 4th International Aviation Security Technology
Symposium, Washington, D.C., USA, November 27 - December 1, 2006,
pp. 303–308.
[7] A. Schwaninger, D. Hardmeier, and F. Hofer, “Aviation security
screeners visual abilities & visual knowledge measurement,” in IEEE
Aerospace and Electronic Systems, vol. 20(6), 2005, pp. 29–35.
[8] J. Bortz, Statistik fu¨r Human- und Sozialwissenschaftler, 6th ed. Wien,
New York: Springer, 2004.
[9] H. Coolican, Research Methods and Statistics in Psychology, 4th ed.
London: Hooder & Stoughton, 2004.
[10] A. Schwaninger, “Evaluation and selection of airport security screeners,”
in AIRPORT, vol. 2, 2003b.
[11] J. Cohen, Statistical Power Analysis for the Behavioral Sciences, 2nd ed.
Hillsdale, New Jersey: Lawrence Earlbaum Associates, 1988.
[12] F. Hofer and A. Schwaninger, “Using threat image projection data for
assessing individual screener performance,” in AIRPORT, vol. 2, 2003b,
pp. 417–426.
[13] ——, “Reliable and valid measures of threat detection performance in
x-ray screening,” in IEEE ICCST Proceedings, vol. 38, 2004, pp. 303–
308.
[14] D. M. Green and J. A. Swets, Signal Detection Theory and Psy-
chophysics. New York: Wiley, 1966.
APPENDIX
For more detailed information on the concepts and math-
ematical models of the image based factors as well as for
examples for a better understanding of the formulae refer to the
on-line technical documentation by Bolfing, & Schwaninger
(2007) made available at:
http://www.psychologie.uzh.ch/vicoreg/publications/index byarea.htm
FTI View Difficulty
The general FTI view difficulty equation 1 describes a
slight modification of the mean of the inverted detection
performance value (DetPerf) over all items (indices NOV )
containing the same FTI object (indices O) in the same view
(subindices V ) as does the item in question. Inverted means
here, that the measured detection performance is subtracted
from the theoretical maximum detection performance. The
slight modification refers to the exclusion of the item in
question from averaging.
FtiVDOV j =
NOV
∑
i=1,j =i
(max(DetPerf) − DetPerfOV i)
NOV − 1
(1)
For analyzing TIP data the inverted detection performance
is the miss rate because usually only bag images containing
threat items are recorded. If a large TIP data set is used, the
exclusion of the item in question from the averaging can be
abandoned due to its very small weight.
FtiVDOV =
NOV
∑
i=1
MissRateOV i
NOV
(2)
Superposition
Superposition equals the inverted Euclidean distance be-
tween the SN images (signal plus noise or threat) and N images
(noise or non-threat images) regarding pixel intensity values.
SP = C −
√
∑
x,y
(
ISN (x, y) − IN (x, y)
)
2
(3)
Clutter
This image based factor is designed to express bag item
properties like textural unsteadiness, disarrangement, chaos or
just clutter.
The method used in this study is based on the assumption, that
such textural unsteadiness can be described mathematically in
terms of the amount of high frequency regions.
Equation 4 represents a convolution of the empty bag image
(N for noise) with the convolution kernel derived from a high-
pass filter in the Fourier space. IN denotes the pixel intensities
of the harmless bag image. F−1 denotes the inverse Fourier
transformation. hp(fx, fy) represents a high-pass filter in the
Fourier space. BS represents bag size (see equation 6). Cut-
off frequency f and transition d (the filter’s order) were set to
f = 0.03 and d = 11. The pixel summation on the high-pass
filtered image was restricted to the bag’s area.
CL =
∑
x,y Ihp(x, y)
BS
(4)
where Ihp(x, y) = IN ∗ F−1(hp(fx, fy))
= F−1(F(IN · hp(fx, fy))
and hp(fx, fy) = 1 −
1
1 +
(
√
f2x+f
2
y
f
)d
Opacity
Opacity reflects the extent to which x-rays are able to
penetrate objects in a bag. These attributes are represented
in x-ray images as different degrees of luminosity. Equation 5
simply implements the number of pixels being darker than a
certain threshold (e.g. 64) in the numerator relative to the bag’s
overall size (denominator). BS represents the formula of the
image based factor bag size (see equation 6).
OP =
∑
x,y
(
IN (x, y) < 64
)
BS
(5)
Bag Size
The bag size formula below is applicable to grayscale
images represented by pixel luminosity values between 0
(black) and 255 (white). All pixels with luminosity lower than
254 (near white) are counted and summed up.
BS =
∑
x,y
(IN (x, y) < 254) (6)
THIRD INTERNATIONAL CONFERENCE ON RESEARCH IN AIR TRANSPORTATION                 FAIRFAX, VA, JUNE 1-4 2008
ISBN: 978-0-615-20720-9324