Sens Imaging (2011) 12:15–33

DOI 10.1007/s11220-011-0059-0

ORIGINAL PAPER

Comparative Data Processing Approaches for Thermal

Wave Imaging Techniques for Non-Destructive Testing

Venkata Subbarao Ghali • Ravibabu Mulaveesala

Received: 15 April 2010 / Published online: 19 March 2011

Ó Springer Science+Business Media, LLC 2011

Abstract Thermal non destructive testing is a whole-field, non-contact and non-

destructive inspection method used to reveal the surface or subsurface anomalies in
a test sample. This paper describes a novel modelling and simulation technique of a
three dimensional pulse compression method for non-stationary thermal imaging.
This method requires much lower peak power heat sources than the widely used
conventional pulsed thermographic methods (PT and PPT) and requires less time
than sinusoidal modulated Lock-in Thermography (LT). In addition, simulation
results obtained with the proposed techniques are compared with the conventional
phase-based thermal imaging techniques (PPT and LT).

Keywords Infrared thermography  Thermal waves  Non stationary signals 

Phase images  Pulse compression  Correlation  Non-destructive testing

1 Introduction

Thermal non-destructive testing (TNDT) involves mapping the temperature

distribution over a test object for the purpose of detecting surface and subsurface
defects. It is a whole field, non-contact and a non-destructive testing method for
defect detection. Since most solids conduct heat, TNDT has potential for wide use in
defect detection in a variety of materials such as metals, composites and
semiconductors [1–14]. TNDT has numerous applications in aeronautical, space,
electrical, electronic and mechanical industries. Of the various possibilities for
TNDT implementation, Infra Red Thermography (IRT) has gained wide acceptance

V. S. Ghali  R. Mulaveesala (&)

Electronics and Communication Engineering Research Group, PDPM-Indian Institute of
Information Technology Design and Manufacturing, Dumna Airport Road, Khamaria (p.o.),
Jabalpur 482005, India
e-mail: ravibabucareiitd@yahoo.co.in

123
16 Sens Imaging (2011) 12:15–33

in non-destructive testing and evaluation (NDT & E) [1]. Various methods and
techniques have further been developed throughout the world to improve and widen
the use of IRT for non-destructive characterization [4–23]. TNDT can be carried out
either in active mode or in passive mode [1].
Passive thermography [1] involves mapping the temperature profile of a sample
surface in the absence of any external heat stimulus. This approach may not provide
sufficient temperature contrast over the defective and non-defective regions of the
test specimen, especially for defects laying deep inside. In order to reveal these
deeper defects with a high contrast, active thermography is used. This requires an
external thermal stimulus to the inspected specimen in order to obtain significant
temperature differences which are indicative of the presence of subsurface defects.
In the active approach [1], an external heat stimulus is provided to the test
sample. Known characteristics of the external thermal stimulation applied onto the
specimen (i.e. nature of excitation, its time duration and its bandwidth etc.),
facilitates the qualitative and quantitative characterization of sub-surface defects.
The present modelling and simulation work is mainly focused on the active
approach.
During the last two decades, intensive modelling, simulation and experimentation
work has been carried out by various researchers throughout the world to introduce
new thermal non-destructive testing methods [1] to overcome limitations in the
existing methods. The most popular TNDT methods are: Pulsed Thermography (PT)
[1–14], Lock-in Thermography (LT) [15–18] and Pulsed Phase Thermography
(PPT) [19, 21]. The choice of any of the above mentioned thermographic methods
for NDT & E depends on the intended application, thermal properties of material to
be tested, defect location and its thickness. Each method has its own advantages and
limitations.
In PT [6–9], the examined sample is warmed up with a short-duration, high peak-
power pulse, and the resultant surface thermal response is recorded. The resultant
sequence of images recorded contains information about defects in the material at
different depths. In practice, this technique requires high peak power heat sources
[13, 14] and has the inherent drawback of being sensitive to surface emissivity
variations and non-uniform heating on the surface of test sample. In general, pulsed
thermography systems are currently preferred for industrial applications. Though
image processing techniques do help to improve the capability of the pulsed
thermographic techniques for subsurface defect detection with improved resolution
and sensitivity, the requirement of high peak power heat sources still remains a
major drawback of PT [13, 14]. However, wave thermography does have some
advantages over PT. In contrast to pulsed thermography, lock-in thermography (LT)
[15–17] is based on thermal waves generated inside the specimen under study.
Mono-frequency sinusoidal thermal excitation at an angular frequency of x,
introduces highly attenuated, dispersive thermal waves of the same frequency (x/
2p) inside the test specimen. The excitation frequency in LT is chosen depending on
the sample thermal characteristics and its geometrical dimensions. The lower the
frequency of the thermal waves, the lower is the velocity in the test specimen, and
the deeper is its penetration into the test specimen. Information about the phase and
magnitude of the reflected thermal wave is derived from the acquired image

123
Sens Imaging (2011) 12:15–33 17

sequence in the stationary regime of the heat cycle. Phase images have several
advantages including being less sensitive to non-uniform illumination of heat
sources and variations of surface emissivity over the sample [15, 16]. Even though it
requires a longer exposure time, another advantage of LT is the relatively small
increase in temperature of the object, which makes it preferable when the testing
specimens sensitive to temperature variations. In addition, phase images are capable
of probing deeper defects compared to the magnitude images [15–18]. Due to its
mono frequency excitation, the generated thermal wavelength inside the test sample
is fixed, leading to a fixed depth resolution. Therefore, in order to get good
resolution for various defects located at different depths inside the test specimen, it
is necessary to repeat LT with different excitation frequencies [22–29].
A research at the University of Laval has proposed a technique called pulsed
phase thermography (PPT) [19, 20] which has some of the advantages of both
conventional PT and modulated LT. The experimental arrangement of PPT is
similar to PT, but the extraction of the various frequency components in the
captured infrared image sequence is performed by Fourier transform (FT) on each
pixel of the thermogram sequence. The phase images obtained from the Fourier
transform in PPT retain all of the advantages of the phase images as obtained in LT,
(i.e. less sensitive to surface in-homogeneous emissivity and illumination
variations). Theoretically, the short duration excitation pulse in PPT does launch
a large number of frequency components into the test samples, but the higher order
frequency components may not have sufficient energy to cause a thermal wave to
propagate deep into the sample. In order to detect deeper subsurface defects in test
sample, PT and PPT require high peak power heat sources [25–28], which may
damage the surface of the test sample.
In order to overcome these limitations of LT and pulse based thermographic
methods, it is necessary to send a desired band of frequencies with significant
magnitude into the test sample [22, 25]. It is preferable that this be done in a single
run without repeating the experiment at different frequency bands. The thermal
excitation should be intense enough to generate thermal waves of appreciable
magnitude for the desired band of frequencies to be launched into the specimen
[28].
In light of the above described limitations of the widely used conventional
methods, frequency modulated thermal wave imaging methods [22, 26, 27] have
been recently introduced in order to improve depth resolution of detecting defects
lying at different depths, in less time compared to LT and with lower peak power
heat sources compared to pulsed thermographic methods.
In frequency modulated thermal wave imaging (FMTWI) [22], the incident heat
flux is varied by driving the heat sources with a linear frequency modulated signal
(in up chirp form), which causes a similar frequency modulated surface heating over
the sample. This helps to inject the desired band of frequencies with significant
magnitude into the test sample which improves the test resolution [24, 26, 28].
In contrast to FMTWI, modulation of the heat sources is much easier in digitized
frequency modulated thermal wave imaging (DFMTWI) [27]. Furthermore, in
DFMTWI we can inject more energy into the sample by probing higher frequency

123
18 Sens Imaging (2011) 12:15–33

harmonics along with the desired band of frequencies which may improve the depth
resolution for near surface defects [22–27].
This paper presents a comparative study of these widely used conventional
methods like PPT and LT with the recently proposed frequency modulated thermal
wave imaging techniques (FMTWI and DFMTWI) using finite element modelling
and simulations.

2 Theory

In this paper, an investigation has been carried out to determine the influence of
noise on correlation-based processing for FMTWI, as compared to phase-based
processing. In addition, FMTWI has been compared with other phase-based
contemporary thermal imaging methods. Details of the FMTWI and pulse
compression are introduced in the following sections.

2.1 Physics of Frequency Modulated Thermal Wave Imaging

In FMTWI, a frequency modulated heat flux imposed over the sample generates
thermal waves as a result of thermal oscillations at the surface. These generated
thermal waves diffuse into the interior portions of the substances under test by
producing similar time-varying temperature distributions on the surface. Thermal
diffusion in substances can be described by the heat conduction equation given by
o2 T ðX; tÞ 1 oT
¼ ð1Þ
oX 2 a ot
where a the thermal diffusion coefficient and T is the instantaneous temperature
at depth X in the direction of heat flow (perpendicular to the surface). Solution
of the above equation for FMTWI under the boundary conditions that (a) the
sample is excited by chirped heat flux at X = 0 for t [ 0 and a similar tem-
perature response is expected at X = 0, and that (b) at X = ? temperature is
zero, results in [22]
pffiffiffiffiffiffiffiffiffiffiffiffi  pffiffiffiffiffiffiffiffiffiffiffiffi  2 
T ðX; tÞ ¼ T0 eX aðf þ s Þ ej X aðf þ s Þ2p ftþ 2s
p Bt p Bt Bt
ð2Þ
where B/s is the frequency sweep rate of the chirp and s is duration of excitation.
Penetrating thermal waves are attenuated by the sample material. The depth at
which the energy of the wave attenuates to 1/e times of its surface value is called
thermal diffusion length, which plays a vital role in thermography. Thermal
diffusion length of this frequency modulated thermal wave is given by
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
a
lfm ¼ ð3Þ
pðf þ Bt=sÞ
Dependence of thermal diffusion length on frequency and bandwidth of the
applied signal can be used to improve the depth resolution by probing different
depths using a range of frequencies in a single experimentation cycle of FMTWI.

123
Sens Imaging (2011) 12:15–33 19

2.2 Pulse Compression with Frequency Modulated Excitations

Pulse compression with non-stationary signals is a signal processing technique

prevalent in radar designed to enhance target detection sensitivity and resolution. It
allows the transmission of a low peak-power, long duration (s) modulated wave
having a bandwidth B. This provides detection and range resolution comparable to
that achieved by short duration, high peak-power pulse methods (Pulsed radars),
even in the presence of interfering signals (noise).
Before applying the principle of pulse compression in thermal non-destructive
testing, the analogy between ‘air-borne target’ detection using a (pulse compression)
radar, and ‘subsurface defect detection in a sample’ using thermal waves, needs to
be studied and their similarities and differences should be understood. In radar, a
chirped pulse is transmitted and reflected back after hitting the target. The reflected
wave is pulse compressed by means of a convolution filter. In IRNDT, the time-
varying heating of the sample surface results in the generation of a thermal wave
that propagates and reflects from the boundary of any sub-surface defect. The
resultant temperature response at the surface is not only attenuated, but also delayed
depending on the depth of the reflecting boundary of the subsurface anomaly. In
radar, the time delay between the reflected echo and the excitation can be estimated
using the delay between the peaks of the compression profiles, and thus the target
range can be estimated.
In the present context, the temperature profiles of different locations can be cross
correlated with the chosen reference profile and depth information can be assessed
by means of the compression peak delays.

2.2.1 Pulse Compression Methods

Pulse compression is achieved either using convolution (radar) or using correlation

(in the present context) depending on the relationship between the reference
(impulse response of the filter) and the signal to be compressed. This section
introduces both the types of pulse compression approaches and also illustrates the
need for correlation-based compression for IRNDT.
Figure 1 illustrates the principle of pulse compression using convolution as is
widely used in radar applications. Let s(t) represent reflected signal from a target,
being essentially same as the transmitted signal except for a reduction in
amplitude and a finite delay. Consider a filter (matched filter), whose time
response is matched to the transmitted waveform, i.e. its impulse response h0 (t) is
the time inverse of s(t). Thus, for example, if s(t) is a linear up-chirp (frequency
increasing linearly with time) waveform, then its matched filter will be a linear
down chirp (frequency decreasing linearly with time). When the signal s(t) goes
through its matched filter, the output g0 (s) (Fig. 1), will be the convolution of the
signal s(t) and its time inverted version h0 (t). This will result in a narrow peak
called the compressed pulse, and can be represented as the convolution of
s(t) with h0 (t) as follows.

123
20 Sens Imaging (2011) 12:15–33

Fig. 1 Principle of convolution based pulse compression for a chirped sine wave s(t)

Z1
0
0
g ðsÞ ¼ sðtÞ h ðs  tÞ dt ð4Þ
1

In order to achieve matched filtering using convolution, the reference signal is

time inverted before the convolution is performed. During the convolution process,
the reference signal undergoes another time inversion and then the sum of the
products of the time-shifted sequences is computed at different time delay instants.
Figure 2 illustrates the principle of correlation-based pulse compression. In this
approach, the received signal is cross correlated with the impulse response h(t) of
the matched filter (which is similar to the received signal except for a finite
attenuation and delay, rather than the h0 (t) used in convolution), without any time
inversion. The correlation of the reference signal (impulse response) and the
reflected signal from the object can be represented as
Z1
0
gð s Þ ¼ sðtÞh ðs þ tÞdt ð5Þ
1

As the result, the long duration (s s) signal s(t) is compressed to a duration

(1/B), governed by the bandwidth B (Hz) of the transmitted/received waveform.
The ratio of the time durations of transmitted and compressed signals is called
the compression ratio and equals (s B) (the time-bandwidth product. This process
compresses each elemental signal into a narrow spike. The signals, now resolved,
constitute a fine resolution in time representing the reflected object position.
If the excitation signal is symmetric, then flipping it back to front does not
change it. So convolution and correlation are the same for measuring the time
delay.

123
Sens Imaging (2011) 12:15–33 21

Fig. 2 Principle of correlation-based pulse compression for a chirped sine wave s(t)

2.2.2 Correlation-Based Pulse Compression for FMTWI

In FMTWI, a frequency modulated (chirp) heat flux imposed over the surface of the
test object produces similar thermal oscillations and generates thermal waves over
the surface. These thermal waves diffuse into the object and reflect back at the
boundary of the anomalies due to the different thermal properties of the anomaly
and results in a differential temperature contrast over the surface. The resultant
thermal response on the surface at the location of defects (where the defects exist
beneath the surface) differs not only in magnitude but also more importantly in
delay depending on the depth of the defects. In order to assess the delay and to
detect the subsurface anomalies, we have adopted a cross correlation-based pulse
compression technique between the resultant thermal (s(t)) at different positions
with that of a similar chosen reference profile (h(t)) (profile of the non defective
location) using Eq. 5. Compressed profiles so obtained concentrate most of the
energy into a narrow main lobe of the pseudo pulse (as shown in Fig. 2) about a
group delayed instant corresponding to the depth of the defect. Pulse compression
also improves the signal to noise ratio (SNR) given by [27–29]
 
S=
N dB ¼ 10 log10 ðsBÞ ð6Þ

where B is the bandwidth and s is the duration of the signal.

2.2.3 Pulse Compression of Thermal Profiles Contaminated with Random Noise

In order to analyse the noise effect on correlation, consider a signal contaminated

with noise such as

123
22 Sens Imaging (2011) 12:15–33

X ðtÞ ¼ sðtÞ þ nðtÞ ð7Þ

where s(t) is thermal profile, n(t) is the contaminating noise content.
If this signal is fed to the filter with impulse response h(t), the resulting output is
y ð t Þ ¼ ð s ð t Þ þ n ð t Þ Þ  h0 ð t Þ ¼ s ð t Þ  h 0 ð t Þ þ n ð t Þ  h 0 ð t Þ ð8Þ
0
From the concept of matched filtering h (t) is simply time inverted function of the
given signal s(t)
h0 ðtÞ ¼ k sðtÞ ð9Þ
This can be achieved with correlation, rather than convolution followed by time
inversion.
From Eqs. 8 and 9
yðtÞ ¼ sðtÞ  h0 ðtÞ þ nðtÞ  h0 ðtÞ
¼ sðtÞ  ksðtÞ þ nðtÞ  ksðtÞ
ð10Þ
¼ k sðtÞ  hðtÞ þ knðtÞ  hðtÞ
¼ k Rss ðtÞ þ knðtÞ  sðtÞ
where * represents convolution,  represents correlation, and Rss represents the
auto correlation of the signal. If the noise is not correlated in time with the signal,
the second term tends to zero. In the present context, contaminated noise is different
at different location and therefore is not correlated to the reference signal.

3 Results and Discussion

In order to analyze the effect of noise on the correlation-based processing for

FMTWI and further to compare the effect of noise on conventional infrared imaging
methods, finite element analysis (FEA) based simulations have been carried out on a
plain carbon steel sample with voids of thickness 0.05 cm at various depths as
shown in Fig. 3. The plain carbon steel sample (shown in Fig. 3) has been modelled

Fig. 3 Dimensional layout of the modelled plain carbon steel sample (a) sample with voids of thickness
0.05 cm and diameter 1 cm at depths a = 0.3 cm, b = 0.35 cm, c = 0.4 cm, d = 0.45 cm, e = 0.5 cm,
f = 0.55 cm, g = 0.6 cm, h = 0.65 cm. b Side view (along thickness) of the modelled sample

123
Sens Imaging (2011) 12:15–33 23

with a fine mesh using 3D tetrahedral elements. The thermal FEA was carried out by
imposing a suitable heat flux over the surface of the test object according to the
scheme and the surface thermal response has been captured at a frequency of 20 Hz.
The simulations were carried out by loading the sample properties such as: thermal
conductivity 43w/m-K, specific heat 440 J/Kg-K and density 7,800 kg/m3 etc.,
under adiabatic boundary conditions, with the sample at an ambient temperature of
300 K.
Analysis of noise influence on correlation group delays and correlation
processing is discussed in Sect. 3.1 and a comparative study of effect of noise on
detection in different thermal methods has been highlighted in Sect. 3.2

3.1 Effect of Noise on Correlation-Based Processing for FMTWI

In order to understand the effect of random noise on the correlation-based pulse

compression approach, noise has been introduced to the simulations in two stages.
Temporal noise due to photonic interactions in IR systems, background noise, and
thermal noise from excitation sources are the common noise sources encountered in
IRNDT. They were modelled and their influence has been studied in two stages [28,
31–33]. In the first stage, prior to recording the temperature distribution over the
sample, source noise has been modelled as zero mean, variance of 0.02 K white
Gaussian noise [28, 31, 32] of SNR 1 dB and added to the imposed heat flux.
In order to understand the effect of random noise on the correlation-based pulse
compression approach, temporal temperature data at a chosen reference location
over the sample (Fig. 4) is considered. Mean increase of the transient response is
removed from the thermal profile obtained over the sample surface (Fig. 4a) by a
proper fit as illustrated in Fig. 4b. This process is repeated for all the temporal
temperature responses over the sample.
Cross correlation has been carried out between the mean-removed reference
profile and the mean-removed profiles obtained at different defect locations over the
sample are as illustrated in Fig. 5.

Fig. 4 Illustration of the obtained temporal temperature data at a chosen non-defective reference location
over the sample including modelled source noise. a Transient temporal thermal profile. b Its mean-
removed counterpart

123
24 Sens Imaging (2011) 12:15–33

Fig. 5 Illustration of the compressed pulses obtained from the cross correlation of the reference signal
with the temporal temperature distributions at different locations over the sample with source noise.
(Enlarged view of correlation peaks is shown in the insert)

In the second stage, before post processing of the recorded temperature distribution
over the sample, camera noise has been modelled as a white Gaussian noise of 20 dB
with a standard deviation of 0.016 with zero mean and Poisson noise of standard
deviation 4.577, is added to the temporal data (Fig. 6) obtained over the sample.
Then the compression process using the correlation approach is repeated for this
noisy data. Figure 7 shows the correlated compressed peaks from various defects
located at different depths for mean-removed noisy sequences.
Figures 5 and 7 show the cross correlation of the reference signal with the
reflected signal without adding noise and with camera noise, respectively. It is clear
from these figures, that there is no change in correlation peak shift, though there is a
decrease in the correlation coefficient value. This illustrates that the correlation-
based pulse compression technique is well suited to measure the depth of defects
even if the signal is contaminated in random noise. Further comparative analysis of
defect detection has been conducted [34, 35] using image SNRs as shown in
Table 1.

Fig. 6 Illustration of the noisy temporal temperature distribution at a location over the sample. a With
the addition of modelled camera noise to the temporal thermal profile and b its mean-removed counterpart

123
Sens Imaging (2011) 12:15–33 25

Fig. 7 Illustration of the compressed pulses obtained from the cross correlation of the reference signal
with the noisy temporal temperature distributions at different locations over the sample. (Enlarged view
of correlation peaks is shown in the insert)

Table 1 Comparison of obtained SNR values for different thermal wave imaging techniques
Defect SNR of defects (dB)

LT PPT FMTWI DFMTWI

Phase Phase Phase Correlation Phase Correlation

a 47.31 59 40.40 67.8 57.11 66.68

b 41.46 51.2 39.15 62.75 54.85 63.67
c 39.58 47.12 38.86 62.54 53.71 64.09
d 32.77 43.65 38.4 58.98 49.75 60.28
e 23.76 34.82 31.1 52.15 43.38 53.47
f 19.2 29.08 27.56 45.88 38.4 48.32
g -15.7 25.34 25.92 40.48 26.69 41.74
h -45 22.01 22.21 34.02 26.65 39.62

3.2 Effect of Noise on Defect Detection in Various IRNDT Techniques

Figure 3 shows dimensional layout of the modelled plain carbon steel sample
having voids of thickness 0.05 cm and diameter 1 cm. FEA simulations were
performed on the sample using Solid Works.
Four thermal non-destructive testing methods: LT, PPT, FMTWI and (its
digitized version, DFMTWI) were simulated on the modelled sample and the results
were compared. In order to find the detection capabilities under noisy conditions,
noise has been added at the corresponding stages as explained in the previous
section and the data is further processed.
Two types of data processing approaches have been implemented on the noisy
acquired surface temperature data of the simulated sample. First, for lock-in

123
26 Sens Imaging (2011) 12:15–33

thermography and pulsed phase thermography, the conventional phase approach has
been adopted, whereas for the frequency modulated thermal wave imaging methods
(linear frequency modulated thermal wave imaging and its digitized version,
digitized frequency modulated thermal wave imaging) in addition to the conven-
tional phase approach, a recently proposed pulse compression approach [22–26] is
also introduced and compared with the phase approach.
In the phase approach, phase information was extracted using the one-dimensional
Fast Fourier Transform (FFT) on each pixel of the thermogram sequence using
MATLAB [19]. This was repeated for all the pixels in the field of view, in order to
obtain the phase images for various frequencies from the generated image sequence.
The correlation-based pulse compression approach has been implemented on the
extracted noisy temperature profiles over the sample for the given frequency
modulated active heating. In this approach, the temperature response over the sample
at a given location over the non-defective region has been considered as the reference.
Correlations of the temperature responses at different locations over the sample with
respect to the chosen reference (over the non-defective region) were obtained [24].
Defect visibility has been quantified in terms of the defect SNR given by
 
Mean of the defective area  Mean of the non defective area
SNR ¼ 20 log
Standard deviation of the non defective region
ð11Þ
Defect SNRs have been computed using the above formula for images obtained
with different schemes and analysis methods as presented in Table 1.

3.2.1 Lock-in Thermography

LT [15–18] is implemented on the modelled specimen (Fig. 3) by heating the

sample surface using a sinusoidal heat flux (10.5 KW/m2) at frequency of 0.04 Hz
for a duration of 50 s in two cycles with an additive white Gaussian noise with SNR
of 1 dB. The temperature data is acquired during the active heating. Extraction of
phase information is performed. Figure 8 shows the phase image, in which defects a
to e are clearly visible under noisy background and defects g and h are contaminated
by the noise. Due to its larger signal strength, phase contrast provided by the
shallower defects is higher even under noisy conditions. It is observed from SNRs of

Fig. 8 Phase image obtained

with lock-in thermography at an
excitation of frequency of
0.04 Hz

123
Sens Imaging (2011) 12:15–33 27

Table 1 that, detection ability of the lock in for deeper defects is observed to be
severely affected by the noise.

3.2.2 Pulsed Phase Thermography

PPT [20] is implemented on the modelled sample by imposing pulse shaped heat
flux (amplitude 105 KW/m2, on duration 5 s with 50 s period) with an additive
white Gaussian noise of SNR of 1 dB over the specimen for a on duration of 5 s and
the temperature data is acquired for 50 s duration. Extraction of phase information
of various frequency components is performed with the one dimensional Fast
Fourier Transform on each pixel of the thermogram sequence. Contributions from
multiple frequencies of pulsed excitation, resolution and detection of the defects in
this approach are observed to be better than lock-in. It has been found from SNRs of
Table 1. It provides better SNR especially for shallowest defect (a) than any other
scheme, still SNRs of deeper defects resembles the effect of noise.

3.2.3 Linear Frequency Modulated Thermal Wave Imaging

Linear frequency modulated (sine chirp) temperature distribution [22] is obtained by

imposing a similar frequency modulated incident heat flux (10 KW/m2) from 0.05
to 0.5 Hz over the modelled sample for a duration of 50 s. Both phase and
correlation data processing approaches have been implemented.
Phase images corresponding to various frequencies are extracted from the
generated image sequence. Figure 10 shows the phase image obtained for a frequency
of 0.2 Hz.
Correlation-based pulse compression approach [25] has been implemented on the
extracted temperature profiles of the sample surface. Figure 11 shows the group
delayed (0.75 s) pulse compressed image obtained by correlation approach. The
correlation peaks clearly illustrate the capability to detect deeper defects in the
simulated sample (Fig. 3) with enough depth resolution.
It is clear from the Fig. 11, that all the defects are clearly visible with SNRs as
given in Table 1, where as in Fig. 10, detected defects have relatively low SNRs as
image is contaminated by the noise. In Fig. 11, variations of contrast exhibited by
the defects (decrease in correlation coefficients with increasing depth) clearly
represent their relative depth.

3.2.4 Digitized Frequency Modulated Thermal Wave Imaging

Digitized linear frequency modulated (square chirp) thermal wave imaging is

implemented by imposing a digitized version of linear frequency modulated
incident heat flux (10 KW/m2) from 0.05 to 0.5 Hz over the modelled specimen for
a duration of 50 s with an additive white Gaussian noise with an SNR of 1 dB. Both
phase and correlation data processing approaches have been adopted on the noisy
temperature distribution. Phase images corresponding to various frequencies are
extracted from the generated image sequence. Figure 12 shows the phase image
obtained for a frequency of 0.2 Hz.

123
28 Sens Imaging (2011) 12:15–33

Correlation-based pulse compression approach has been implemented on the

temperature responses at different locations over the sample surface with respect to
the chosen reference (non-defective region over the sample). Figure 13 shows the
group delayed (0.75 s) pulse compressed image obtained by the correlation
approach. The correlation peaks clearly illustrate the capability to detect deeper
defects in the simulated sample (Fig. 3) with adequate depth resolution.
It is clear from the Fig. 13 that in the correlation image, all the defects are visible
with good contrast and better SNR where as in the phase approach, (Fig. 12) all the
defects are detectable with reduced SNR. It is observed that the correlation image is
providing better depth resolution and visibility than the phase images, due to the
improved SNR of compressed responses from Eq. 6.

3.2.5 Comparison of Thermal Wave Imaging Methods

Simulation results shown in figures (Figs. 8, 9, 10, 11, 12, 13), indicate that the
pulse compression approach to non stationary thermal excitations exhibits better
contrast [30] in defect detection (Figs. 11, 13) than the conventional phase approach
(Figs. 10, 12). It is observed from Table 1 that the pulse compression approach
enhances SNRs by approximately 10–20 dB over the phase approach. Even for the
phase approach, phase images obtained with non-stationary excitation methods
show better shape preserving capability than the conventional lock-in thermography
phase image. Furthermore, FMTWI and DFMTWI methods can be performed at
much low peak power excitations than PPT. Figures 11 and 13 show that the pulse
compression images obtained from the correlation approach have better depth
resolution for deeper defects than the conventional phase images (LT and PPT etc.)
(Figs. 8, 9, 10, 12). Table 1 shows the obtained SNR values [31, 32] in dB for the
defects using LT, PPT, FMTWI and DFMTWI.
The larger temperature rise of the sample due to applied heat flux may cause
damage to the sample. Schemes which are able to provide adequate detail are better
suited for thermal NDT applications.

Fig. 9 Phase image obtained

with pulsed phase thermography
at frequency of 0.1 Hz

123
Sens Imaging (2011) 12:15–33 29

Fig. 10 Phase image obtained

with frequency modulated
thermal wave imaging at
frequency of 0.2 Hz

Fig. 11 Pulse compressed

image obtained with frequency
modulated thermal wave
imaging acquired at 0.75th
second

Fig. 12 Phase image obtained

with digitized frequency
modulated thermal wave
imaging at frequency of 0.2 Hz

Fig. 13 Pulse compressed

image obtained with digitized
frequency modulated thermal
wave imaging acquired at 1.25th
second

123
30 Sens Imaging (2011) 12:15–33

Figure 14 represents the typical temperature variations obtained with different

schemes (Fig. 14a) and their mean-removed temperature profiles (Fig. 14b). It is
observed that, even though the temperature rises over the sample in FMTWI and
DFMTWI are relatively lower compared to the remaining schemes, they still
provide the best performance as observed in Figs. 8, 9, 10, 11, 12 to 13.

3.2.6 Defect Depth Quantification Using Correlation Peak Delays

Matched filtering based pulse compression using the correlation approach concen-
trates the applied signal energy in the form of a sinc pulse. Figures 5 and 7 illustrate
the noise influence and the shift invariance of the correlation peaks of the
temperature profiles collected from the centre points of the anomalies at different

Fig. 14 Illustration of temperature profiles of LT, FMTWI, DFMTWI and PT excitation schemes.
a Transient thermal profiles. b Mean-removed counterparts of profiles shown in a

Fig. 15 Illustration of depth profilometric delay measurement of subsurface voids in modelled plain
carbon steel sample (Fig. 3) with frequency modulated thermal wave imaging and its digitized version

123
Sens Imaging (2011) 12:15–33 31

depths. These compressed profiles are centered about their appropriate group delays
which correspond to the depths of the defects.
Figure 15 illustrates the depth dependence of the group delay provided by the
cross correlation of thermal profiles (shown in Figs. 5, 7) of the defects with respect
to the sound surface reference profile. The cross correlation peak delay (with respect
to the peak of auto correlation of the reference) is a measure of the defect depth. The
correlation peaks delays (Figs. 5, 7) were measured and empirically fitted to the
equation
s ¼ aebx þ cedx ð12Þ
where x is the depth of the defect and a, b, c and d are 33.258 9 109, -66.78, 23.77
and -7.16 for FMTWI and 3.896 9 106, -67.04, 24.59, -7.29 for DFMTWI,
respectively. Non linear delay contributions to thermal responses obtained from
different sample depths provide non linear peak delays observed in the cross
correlation profiles of the corresponding thermal responses. DFMTWI, with its
relatively larger bandwidth, provides higher peak delay resolution than FMTWI. It
has been observed that, even for the same depth, the delays obtained using
DFMTWI have significantly better resolution than those obtained using FMTWI.

4 Conclusion

Simulations of a modelled steel sample (finite-element analysis based model) for

sub-surface defect detection with different types of thermal excitations were
performed. Comparison has been made with the widely used conventional phase-
based thermal wave imaging methods such as LT, PPT and the recently proposed
non stationary thermal excitations methods such as FMTWI and DFMTWI. Results
showed that non stationary thermal excitations with pulse compression approach
provide better detection capabilities over the conventional thermal imaging
techniques for sub surface defect detection even under noisy conditions.

References

1. Maldague, X. (2001). Theory and practice of infrared thermography for nondestructive testing.
Hoboken: John Wiley-Interscience.
2. Balageas, D. L., Deom, A. A., & Boscher, D. M. (1987). Characterization and nondestructive testing
of carbon epoxy composites by a pulsed photo thermal method. Materials Evaluation, 45, 461–465.
3. Balageas, D. L., Krapez, J. C., & Cielo, P. (1986). Pulsed photothermal modeling of layered
materials. Journal of Applied Physics, 59, 348–357.
4. Shepard, S., Chaudhry, B., Predmesky, R., & Zaluzec, M. (1998). Pulsed thermographic inspection of
spot welds. Proceedings of SPIE, 3361, 320–324.
5. Tam, A. C., & Sullivan, B. (1983). Remote sensing applications of pulsed photothermal radiometry.
Applied Physics Letters, 43, 333.
6. Avdelidis, N. P., & Almond, D. P. (2004). Through skin sensing assessment of aircraft structures
using pulsed thermography. Journal of NDT & E International, 37(5), 353–359.
7. Shepard, S. M. (2001). Advances in pulsed thermography. Proceedings of SPIE, 4360, 511–515.

123
32 Sens Imaging (2011) 12:15–33

8. Rosencwaig, A., & Gersho, A. (1976). Theory of the photoacoustic effect in solids. Journal of
Applied Physics, 47, 64–69.
9. Cowell, S. D., Burleigh, D. D., & Murray, T. J. (1989). Flash lamp heat flux requirements for
thermographic inspection of fiber composite laminates. Proceedings of SPIE, 1094, 182–187.
10. Almond, D. P., & Lau, S. K. (1994). Defect sizing by transient thermography. I: An analytical
treatment. Journal of Physics D: Applied Physics, 27, 1063–1069.
11. Saintey, M. B., & Almond, D. P. (1995). Defect sizing by transient thermography. II: A numerical
treatment. Journal of Physics D: Applied Physics, 28, 2539–2546.
12. Avdelidis, N. P., Hawtin, B. C., & Almond, D. P. (2003). Transient thermography in the assessment
of defects of aircrafts composites. NDT and E International, 36(6), 433–439.
13. Ringermacher, H. I., Howard, D. R., & Filkins, R. J. (2004). Flash-quenching for high resolution
thermal depth imaging. AIP Proceedings, 700(1), 477–481.
14. Ringermacher, H. I., Howard, D. R., & Knight, B. (2006). Thermal imaging at general electric. AIP
Proceedings, 820(1), 523–528.
15. Busse, G., Wu, D., & Karpen, W. (1992). Thermal wave imaging with phase sensitive modulated
thermography. Journal of Applied Physics, 71, 3962–3965.
16. Vavilov, V., & Marinetti, S. (1999). Pulsed phase thermography and fourier-analysis thermal
tomography. Russian Journal of Nondestructive Testing, 35(2), 134–145.
17. Busse, G., & Eyerer, P. (1983). Thermal wave remote and nondestructive inspection of polymers.
Applied Physics Letters, 43, 355–357.
18. Choi, M., Kang, K., Park, J., Kim, W., & Kim, K. (2008). Quantitative determination of a subsurface
defect of reference specimen by lock-in infrared thermography. NDT and E International, 41(2),
119–124.
19. Ibarra-Castanedo, C., Avdelidis, N. P., & Maldague, X. (2005). Qualitative and quantitative
assessment of steel plates using pulsed phase thermography. Materials Evaluation, 63(11),
1128–1133.
20. Ibarra-Castanedo, C., González, D., Klein, M., Pilla, M., Vallerand, S., & Maldague, X. (2004).
Infrared image processing and data analysis. Infrared Physics & Technology, 46, 75–83.
21. Cielo, P., Maldague, X., Deom, A. A., & Lewak, R. (1987). Thermographic non destructive eval-
uation of industrial materials and structures. Materials Evaluation, 45, 452–460.
22. Mulaveesala, R., & Tuli, S. (2006). Theory of frequency modulated thermal wave imaging for non-
destructive sub-surface defect detection. Applied Physics Letters, 89, 191913.
23. Mulaveesala, R., Pal, P., & Tuli, S. (2006). Interface study of bonded wafers by digitized linear
frequency modulated thermal wave imaging. Sensors and Actuators A, 128, 209–216.
24. Mulaveesala, R., Vaddi, J. S., & Singh, P. (2008). Pulse compression approach to infrared nonde-
structive characterization. Review of Scientific Instruments, 79, 094901.
25. Tuli, S., & Mulaveesala, R. (2005). Defect detection by pulse compression in frequency modulated
thermal wave imaging. Journal of Quantitative Infrared Thermography, 2(1), 41–54.
26. Mulaveesala, R., & Tuli, S. (2005). Implementation of frequency-modulated thermal wave imaging
for non-destructive sub-surface defect detection. Insight, 47(4), 206–208.
27. Mulaveesala, R., & Tuli, S. (2005). Digitized frequency modulated thermal wave imaging for non-
destructive testing. Materials Evaluation, 63, 1046–1050.
28. Mulaveesala, R. (2006). Frequency modulated thermal wave imaging: theory, modeling, simulation
and applications, Ph.D. Dissertation, Centre for Applied Research in Electronics, Indian Institute of
Technology Delhi, New Delhi, India.
29. Wehner, D. R. (1995). High resolution radar (2nd ed.). London: Artech House Inc.
30. Ghali, V. S., Jonnalagadda, N., & Mulaveesala, R. (2009). Three dimensional pulse compression for
infrared non-destructive testing. IEEE Sensor Journal, 9(7), 832–833.
31. Grinzato, E., Vavilov, V., Bison, P. G., Marinetti, S., & Bressan, C. (1995). Methodology of pro-
cessing experimental data in transient thermal NDT. Proceedings of SPIE, 2473, 167–178.
32. MaIdague, X., Largoutit, Y., & Couturier, J. P. (1998). A study of defect depth using neural networks
in pulsed phase thermography: modeling, noise, experiments. Revue Ge´ne´rale de Thermique, 37,
704–717.
33. Petrovi, V. S., & Xydeas, C. S. (2003). Sensor noise effects on signal-level image fusion perfor-
mance. Information Fusion, 4, 167–183.
34. Omar, M. A., & Zhou, Y. (2008). A quantitative review of three flash thermography processing
routines. Infrared Physics & Technology, 51(4), 300–306.

123
Sens Imaging (2011) 12:15–33 33

35. Lee, D. J., Mitra, S., & Krile, T. F. (1989). Analysis of sequential complex images, using feature
extraction and two-dimensional cepstrum techniques. Journal of the Optical Society of America A, 6,
863–870.

123