12

Fire Technology, 45, 313322, 2009

2008 Springer Science+Business Media, LLC. Manufactured in The United States
DOI: 10.1007/s10694-008-0060-x

Image Frequency Analysis for Testing

of Fire Service Thermal Imaging Cameras
Francine Amon*, National Institute of Standards and Technology, Gaithersburg,
MD, USA
Alfred Ducharme, University of Central Florida, Orlando, FL, USA
Received: 20 September 2007/Accepted: 12 June 2008

Abstract. A growing number of rst responders are purchasing infrared cameras

(thermal imagers). Thermal imagers provide critical information for many reghting
operations: size up, tracking of re growth, location of victims and egress routes,
overhaul, etc. Although thermal imagers have proven to be useful tools for the re
service over the past 1015 years, there is no existing comprehensive performance
standard for thermal imagers used in re service applications. A suite of thermal imager performance metrics and test methods is being developed for possible inclusion in
a proposed standard on thermal imagers for the re service. The following goals
apply to two of the imaging performance metrics in this suite: to explore the feasibility of a new spatial resolution metric and test method, the Random Modulation
Transfer Function (MTF); and to transfer the essential characteristics of complex re
scene images to a thermal target for the eective temperature range test using image
frequency content analysis. Images of humans standing in an upright position at
varying distances from the thermal imager under test and wearing either street clothes
or turnout gear were analyzed for frequency content. These data were then used to
determine the frequency of a bar target for the bench-scale eective temperature
range test.
Keywords: thermal imaging camera, performance metric, test method, image analysis, re service

1. Introduction
Infrared (IR) technology for re ghting applications has matured to the point
that most rst responder organizations in the United States either have purchased
or are considering the purchase of thermal imaging cameras. Thermal imagers can
provide rst responders with critical information to size up a re incident, track
re growth, and to locate victims, other rst responders, and egress routes. While
these devices represent a signicant investment, typically on the order of $10 K
per camera, rst responders have little guidance on instrument performance
beyond manufacturer literature and recommendations from other users [1]. These
issues are further complicated because the demands placed on thermal imagers are
application dependent. The end users may have very dierent ideas about which
imaging properties are most important: sharp image contrast may be sucient for
* Correspondence should be addressed to: Francine Amon, E-mail: francine.amon@nist.gov

314

Fire Technology 2009

some re ghting applications, such as nding the source of a re, but high thermal sensitivity may be required to locate a person or structural component when
ames and water are in the imagers eld of view. Currently, there are no standardized performance guidelines available to aid end users in making purchasing
decisions.
Over the past several years, the Fire Research Division at the National Institute
of Standards and Technology has been developing a suite of imaging performance
metrics and test methods to recommend for inclusion in a national consensusbased standard on thermal imaging cameras used by rst responders [2]. The performance metrics are related directly to the environment in which the imagers are
used and tasks typically performed by rst responders. Measurements of image
contrast, eective temperature range, spatial resolution, image nonuniformity, and
thermal sensitivity are currently included in the proposed standard, along with
other performance tests that address the robustness of the imager, i.e., resistance
to vibration, immersion, impact, heat, etc. The test methods associated with two
of these performance metrics, e.g., spatial resolution and eective temperature
range, may benet from a careful analysis of the frequency content of the images
that appear on the thermal imagers display. As a subset of the overall project,
NIST is working with the College of Optics and Photonics at the University of
Central Florida to investigate two techniques for transferring the essential characteristics of complex images generally seen in the re environment to relatively simple bench top target congurations for use in standardized performance tests.

2. Methodology
A Fourier transform is a mathematical procedure in which the spatial (or temporal) uctuations in contrast in an image (or sequence of images) are converted
from the spatial (or time) domain to the frequency domain. For a two-dimensional image, this transformation is accomplished using the following equation:
F m; n

XN
1 XM
1
2

l
N2

k
M2

f k; le
2pi M N
mk

nl

where F is the Fourier transformed image, m and n are the dependent variables in
the frequency domain, in this case contrast magnitude and phase, and k and l are
the respective quantities in the spatial domain, in this case pixel intensity and
position [3]. We shall use Fourier transforms in two ways in this discussion: rst,
as an expedient way to assess a thermal imaging cameras spatial resolution; and
second, as a way to extract vital spatial frequency information from realistic images to design bench-scale test targets for use in other proposed performance tests.

2.1. Modulation Transfer Function (MTF)

All image forming optical systems suer from the same performance degradation:
as objects get smaller they become less distinct and more dicult to discern. An
objects decrease in size is equivalent to an increase in spatial frequency. A pure

Image Frequency Analysis

315

Figure 1. Blurring effect of increasing bar frequency. The top half of

the figure represents the target and the bottom half represents the
resulting image reconstructed by the thermal imager.

1
0.9

MTF

0.8
0.7

Nyquist

0.6

frequency

0.5
0.4
0.3
0.2
0.1
0
0.00

0.05

0.10

0.15

0.20

Frequency (cyc/mrad)

Figure 2. Typical MTF curve for a firefighter thermal imager having a

320 3 240 pixel detector array.

spatial frequency can be thought of as a periodic change of light and dark features in an image. A plot of these changes in image pixel intensity would be sinusoidal in nature and exhibit a common peak-to-peak distance (wavelength) the
inverse of which is the spatial frequency of the feature. This sinusoidal uctuation
in intensity is known in optics as modulation.
The MTF is a measure of an imaging systems ability to reconstruct an image
from a target over all spatial frequencies, and varies in value from 0 to 1, where 0
indicates no distinguishing characteristics in the image whatsoever and 1 indicates
that the nest details in the target are retained in the reconstructed image. MTF
decreases with increasing frequency due to blur that aects image contrast as
objects become smaller and closer together [4], this blurring eect is illustrated in
Figure 1 using a bar target of increasing frequency. In other words, the MTF is a
measure of an imagers spatial resolution, or ability to see the details in a scene or
target.
The MTF is conventionally normalized to a value of 1 at the ordinate, or zero
frequency, and is usually presented as a curve with MTF plotted as a function of
frequency, an example of which is shown in Figure 2. An MTF curve is usually
carried out to the Nyquist frequency, which is the frequency at which the sampling rate causes signal aliasing and can introduce image artifacts such as moire

316

Fire Technology 2009

Figure 3. Reconstructed images of point, line, and edge targets used

for MTF measurements. The targets have sharply contrasting areas
which are blurred by the thermal imaging system.

patterns [4]. The Nyquist frequency depends on the geometric design of the optics
and detector of the camera under test.
Typically, sharply contrasting adjacent areas such as points, lines, and edges are
imaged, then Fourier transform analysis is used to examine the transition in pixel
intensity, or blur, as a function of distance from the contrast edge. Figure 3 shows
blurred images of typical contrast targets used to measure MTF.
The MTF is regarded as the one of the most valuable gures of merit for optical imaging systems and is being considered as the basis of the spatial resolution
performance metric developed in this work.

2.2. Random MTF Target

The Random MTF target uses random features distributed in a specic pattern to
yield a uniform distribution of spatial frequency power over a range of frequencies that are determined by the pixel size and conguration of the thermal imagers detector. An important advantage of the Random MTF target approach
over other techniques is repeatability. The theory of MTF is inherently based on
linear shift invariance (LSI), also referred to as spatial variance, see Holsts discussion of MTF theory, including detector-target positional eects [5]. LSI implies
that small movements or positional dierences of the object or test pattern with
respect to the imager detector under test do not yield dierent MTF results. Commonly used MTF targets such as points, lines, bars, and knife-edges all have
sharp edges; these methods require that the target must be carefully aligned with
the rows and columns of the detector array of the imager under test in order to
produce repeatable results. The degree of uncertainty in the test results depends in
large part on the detector position and design, a small movement or shift of the
target in these tests can result in signicantly dierent measurement results
because the sharp target edges may partially ll the cameras detector pixels in
non-repeatable ways, although oversampling techniques may be employed to partially compensate for detector-target positional eects. The Random MTF target,
shown in Figure 4, utilizes random features and incorporates an averaging technique during processing, which results in a test method that is immune to shifts of
the target and is in fact LSI.

Image Frequency Analysis

317

Figure 4. Image of a Random MTF target. This target is linear shift

invariant due to the randomized nature of the contrasting areas. All
frequencies of interest are represented equally.

An MTF test using the Random MTF target is performed by imaging a backlit
transparent target. The backlight in this case is an extended area infrared blackbody set to a nominal temperature. For this application, the thermal imaging
cameras operate in the 8 to 14 lm wavelength range, therefore the target substrate
must be transparent in this range. Zinc selenide, which transmits approximately
6570% in this wavelength range, is the material of choice for this project. Chromium is deposited on the substrate to form the random pattern. The image of the
target produced by the thermal imager under test is then processed using Fourier
transform image analysis techniques [6].

2.3. Image Frequency Content

In the development of a consensus standard, as is this case, performance metrics
and test methods are created in a process in which many diverse parties, such as
rst responders, thermal imager manufacturers, and re researchers, all participate. The goal is to form a consensus on test methods that are meaningful to rst
responders, useful to manufacturers, and are also scientically defensible. For
example, in the proposed eective temperature range test, the thermal imager
views a target, shown in Figure 5, in which the temperature of a large area (representing ames and heat) is increased from ambient to 550C while a bar pattern
having constant intermediate temperatures (representing a human) is analyzed for
contrast. The rationale behind this test is to ensure that the thermal imager provides adequate contrast at relatively low temperatures, such as may be needed to
discern a human against a background of similar temperature, in cases where

318

Fire Technology 2009

Tamb = 25 C

Tbar1 = 25
Hot Area
25 C < T2 < 550 C
Tbar2 = 30

Figure 5. Test target for the proposed effective temperature range

test. The hot area temperature ramps from ambient to 550C while all
other temperatures in the field of view are held constant.

there are hot surfaces, perhaps even ames, in the eld of view. It should be noted
that reghter thermal imagers intended to be compliant with the standard will
employ automatic gain, oset, and integration time functions. It is, therefore,
important that the manufacturers settings for these functions lead to useful images in challenging conditions likely to be encountered by reghters.
The frequency of the bar pattern is important in this test because of the wide
variation in thermal imager design. Some thermal imagers may not have sucient
spatial resolution to see the bars clearly, even under optimum thermal conditions.
Spatial resolution is measured in another of the tests included in the recommendations for the standard. In order to avoid biasing this test method toward any particular thermal imager technology, the bar frequency was determined by the needs
of the users. Input from the re service on the type of image scene that best characterizes their most pressing needs in this kind of thermal environment led to the
development of a bar pattern based on frequency analysis of images of people
dressed in turnout gear and in street clothes standing at several distances from the
thermal imager under test. This process yields target design information that
greatly simplies the test method and reduces the uncertainty of test results while
retaining meaningful characteristics from realistic images.

2.4. Bar Frequency Determination

Before discussing the frequency content analysis of complex images of people, it is
instructive to rst demonstrate the process on simple, well-characterized images of
known frequency. In the top of Figure 6, two such images of bars are shown: vertical and horizontal bars at two dierent frequencies. The bottom two images are
the corresponding Fourier transforms in frequency space. All of the information
in the two upper images is contained in the lower two images, but in a dierent
manner.
The bright point in the center of the two lower images has a magnitude equal
to the average pixel intensity value of the original image. The two bright points
on either side of the center indicate three things: the frequency of the bars, by

Image Frequency Analysis

319

Figure 6. The vertical and horizontal bars shown in the images at

the top of the figure are represented in frequency space in the
corresponding images at the bottom of the figure.

their distance from the center point; the bar orientation, by their angular position
relative to the center point; and the sharpness of the contrast, by the magnitude of
their pixel intensity value. The distance of the bright points from the center point
is measured in pixels; the frequency of the bars is in cycles/pixel, where a cycle is
the width of a pair of bars (one black bar and one white bar). The frequency
space points for two-dimensional Fourier transforms are symmetric about the center point, but in this simple example the only contrast dierences in the original
images occur strictly in either the horizontal or vertical directions, therefore the
frequency space points are angularly aligned either vertically or horizontally symmetric about the center point. In both cases, the contrast magnitude, or brightness, of the points on either side of the center point indicate that the contrast
between bars is high (pure white and pure black). There are usually also other
points of lesser magnitude in the frequency space images that are artifacts of harmonics (mathematical echos) arising from very sharp contrast changes in the original image. In Figure 6, the lower frequency vertical bars are represented by
points located at a relatively short distance from and aligned horizontally with the
center point. Conversely, the higher frequency horizontal bars appear in frequency
space as points located at a relatively longer distance from the center point and
aligned vertically with it. Additional information about using Fourier transforms
to convert images into and out of frequency space can be found in [711].
For this work, infrared images of two people, a woman and a man, dressed
in street clothes and in reghter turnout gear were collected and analyzed at

320

Fire Technology 2009

Figure 7. Infrared images of a person standing 3 m distant from

a thermal imaging camera. He is wearing street clothes (left) and
firefighter turnout gear (right).

distances of 2, 3, 5, and 10 m from the thermal imager, resulting in a total of 16

images for each of 8 thermal imagers. An example of a pair of images taken at
3 m is shown in Figure 7. The backgrounds of the images were removed to avoid
unnecessarily complicating the analysis results.
Given a known dimension of an object that appears in every image, the pixels
in the frequency space images can be equated to actual distances, which are used
to relate the position of the brightest points back to the frequency content of the
original image, measured in cycles/mrad. Since most of the interesting information
in the images was oriented vertically, e.g., neck, torso, arms, a one-dimensional
horizontal Fourier transform was used, with each column summed along the
height (y-axis) of the image. In the plots shown in Figure 8, the magnitudes of the
frequencies obtained for a single thermal imager provide an indication of the frequencies that contribute the most to the overall image.
Hand calculations of the frequencies of dominant features, such as the head,
arms, and torso were performed to check the accuracy of the plots. The magnitude of the relative contrasts of the frequencies in these plots may be somewhat
misleading in that they may include a minor contribution from harmonic eects
coming from lower frequency data. The most dominant features in the images
taken of people in street clothes were the head and arms, while the dominant features in images taken of people in turnout gear were the face (they were not wearing masks) and arms, particularly where body heat seeps into the creases in the
elbow region. Features with sharp contrast and a relatively long y-axis dimension
have a larger magnitude than features that are decient in either or both of these
components. For example, in the street clothes case, the head and arms had similar contrast but the arms have a long y-axis dimension so their magnitude is

Image Frequency Analysis

321

Figure 8. Frequency plots of infrared images of a person dressed in

street clothes (left) and in firefighter turnout gear (right) at four
distances from the thermal imager.

greater than that of the head. In the turnout gear case, the arms were still the
most dominant feature, but not as dramatically as in the street clothes case,
because the contrast magnitude was diminished by shielding eects of turnout
gear.
Results show that, as the person moves further away from the thermal imager
position, she/he appears smaller and thus exhibits relatively higher frequencies.
The spatial resolution of the thermal imager tends to cause the features in the
images to blur together as the distance increases, sometimes working counter-intuitively to enhance the low frequency response as the higher frequencies become
irresolvable. There was little dierence in frequency plots between male and
female targets. The images of people wearing street clothes showed more features
than the corresponding images of people wearing turnout gear and their magnitude was higher. This was attributed to the fact that turnout gear tends to mask
the heat signature of the person wearing it; body heat penetrates street clothes
much more readily. Also, a dierence in the frequency plots was noticed regarding
the location of the peaks, with the peaks being shifted toward higher frequencies
for street clothes. Compared to street clothes (jeans, tee shirt or blouse), turnout
gear is large and bulky and would therefore present a somewhat larger, lower frequency target.
With respect to the optimum bar frequency to use in the eective temperature
range test, it was determined that all the thermal imagers tested could easily discern the most distinctive feature in all of the images (the arms), as well as the
head or face, legs, torso, and shoulders. Most of the features in the turnout gear
images were at or below 0.04 cyc/mrad while some important features in the street
clothes images, particularly at the 10 m distance, were at higher frequencies. It
was therefore decided that an eective temperature range target having a bar frequency of 0.04 cyc/mrad would be more than adequate to represent people wearing turnout gear and would also represent most cases in which victims wearing
street clothes are encountered.

322

Fire Technology 2009

3. Conclusions
The advantages of using these techniques for image analysis are twofold: rst,
inherent diculties related to the exact position of the test target with respect to
the imager detector array can be alleviated if the elements of the target are numerous enough and are randomly arranged; and second, realistic images such as that
of a reghter standing nearby can be replaced with a simple bar pattern having
similar frequencies for bench-scale image performance tests. These two image
analysis techniques exploit these advantages and are being developed for evaluation of rst responder thermal imagers.
In addition to reducing the measurement error associated with LSI, a connection is established between the design of bench-scale test targets used in the eective temperature range test and realistic images that a reghter might expect to
see at an emergency event. Further work is underway in both these areas to assess
the benets of full implementation of these image analysis techniques.

References
1. Amon FK, Bryner NP, Hamins A (June, 2005) Thermal imaging research needs for rst
responders: workshop proceedings. NIST Special Publication, p 1040
2. Proposed NFPA 1801 Standard on Thermal Imagers for the Fire Service, 1801-F09PRE-ROP (September, 2007) National Fire Protection Association
3. Press WH, Flannery BP, Teukolsky SA, Vetterling WT (1986) Numerical recipes: the
art of scientic computing. Cambridge University Press, New York, pp 390407, ISBN
0-521-30811-9
4. Boreman GD (2001) Modulation transfer function in optical and electro-optical
systems. TT52SPIE Press, Bellingham
5. Holst GC (2006) Electro-optical imaging system performance, 4th edn. SPIE Press and
JCD Publishing, Bellingham
6. Ducharme A, Boreman GD, Ducharme AD (1995) Random transparency targets for
modulation transfer function measurement in the visible and infrared regions. Opt Eng
34(3):860868
7. Ballard D, Brown C (1982) Computer vision. Prentice-Hall, Englewood Clis, pp 2430
8. Gonzales R, Woods R (1992) Digital image processing. Addison-Wesley Publishing
Company, Reading, pp 81125
9. Horn B (1986) Robot vision. MIT Press, Cambridge, Chaps 6:7
10. Jain A (1985) Fundamentals of digital image processing. Prentice-Hall, Englewood
Clis, pp 1520
11. Marion A (1991) An introduction to image processing. Chapman and Hall, London,
Chap. 9