Optics and Lasers in Engineering 111 (2018) 8–18

Contents lists available at ScienceDirect

Optics and Lasers in Engineering

journal homepage: www.elsevier.com/locate/optlaseng

High temperature strain measurement method by combining digital image

correlation of laser speckle and improved RANSAC smoothing algorithm
Jinlian Song a, Jianhong Yang a,∗, Fujia Liu a, Kefei Lu b
a
School of Mechanical Engineering, University of Science and Technology Beijing, Beijing100083, China
b
Aerospace Research Institute of Material & Processing Technology, Beijing 100076, China

a r t i c l e i n f o a b s t r a c t

Keywords: Background radiation and airflow disturbance at high temperature reduce the quality of captured images, result-
Digital image correlation (DIC) ing in a lot of noises in the displacement field obtained by digital image correlation (DIC) and decreased accuracy
Laser speckle of strain measurement. A high temperature strain measurement method (IR-DIC) by combining digital image cor-
High temperature strain measurement
relation of laser speckle and Improved Random Sample Consensus (IRANSAC) smoothing algorithm for uniform
Improved Random Sample Consensus
deformation is proposed. In the IR-DIC method, after obtaining the noisy displacement field by the digital image
(IRANSAC)
Smoothing of displacement field correlation of laser speckle, the IRANSAC algorithm is used to smooth the noisy displacement field by removing
noises with an adaptive noise threshold selection for uniform deformation. Also, when fitting the displacement
field, instead of using the minimal data model in the classical RANSAC algorithm, a subset with all the remaining
effective data points after removing noises are used in the IRANSAC algorithm. An experiment on tensile test of
a C/C composite sample at 2000 °C in the inert atmosphere was carried out. The results showed that, the strain
curve calculated by the IR-DIC method is basically consistent with the result simultaneously measured by a high
temperature contact extensometer. And compared with the least square algorithm, the mean deviation between
the strain curve calculated by the IR-DIC method and the strain curve measured by the contact extensometer is
reduced about 13.94%.

1. Introduction
With the development of aerospace engineering, more and more ma-
terials need to serve in high temperature environment. In order to ac-
curately characterize the mechanical behavior of materials at high tem-
peratures, a lot of high temperature strain measurement methods have
been developed in recent decades, among which digital image correla-
tion (DIC) method [1–3] has attracted more and more research inter-
est due to its advantages of non-contact, high accuracy and low sen-
sitivity to complex environment. In 2011, Liu [4] presented a novel
Deformation-Pattern-based Digital Speckle Correlation (DPDSC) method
which improved the accuracy and reproducibility of CTE measurement
of film at high temperature environment. In 2017, Dong [5] developed
an integrated digital image correlation (I-DIC) method to address the
difficult problem of the simultaneous and accurate identification of mul-
tiple thermo-mechanical parameters of superalloy materials in compli-
cated service condition. However, when DIC is used for high temper-
ature strain measurement, there are still many problems to be solved
[6,7], such as the preparation of speckle, the strong background radi-
ation, the airflow disturbance and the strain calculation from displace-
ment field with noises.
Artificial speckle [8] and laser speckle [9,10] are commonly used in
DIC. Traditionally, artificial speckle is formed by randomly spraying the
specimen surface with black or white paint which can be distinguished
from the color of the specimen. In 2017, LePage [11] showed that paint
sequence can be important due to fundamental differences in paint pig-
ments, and provided recommendations for optimum painted patterns
for digital image correlation. Laser speckle is a phenomenon that is ex-
perienced if an optical rough surface is illuminated by a high coherent
laser source. In 2000, Lagattu [12] used the laser speckle method for the
study of composite and polymer mechanical behavior. In 2000, Anwan-
der [13] presented a laser-based noncontacting strain sensor suitable
for temperatures up to 1200 °C, and performed a series of tensile tests
with various materials at room temperature and high temperatures to
test the applicability of the optical strain sensor. Compared with artifi-
cial speckle and laser speckle [14], artificial speckle is easy to prepare
at room temperature, but it is more susceptible to human factors. And
artificial speckle is more advantageous than laser speckle in large de-
formation measurement of the specimen. However, in high temperature
environment, artificial speckle has some shortcomings, such as speckle
falling off, and discoloration. Laser speckle is more suitable for high tem-
perature applications because it does not need to seek special adhesive

∗
Corresponding author.
E-mail address: yangjianhong@me.ustb.edu.cn (J. Yang).

https://doi.org/10.1016/j.optlaseng.2018.07.012
Received 9 May 2018; Received in revised form 21 June 2018; Accepted 24 July 2018
0143-8166/© 2018 Elsevier Ltd. All rights reserved.
J. Song et al.
or speckle paint resistant to high temperature. But laser speckle pattern
will also be influenced by background radiation and airflow disturbance.
In order to reduce negative effects on the accuracy of strain measure-
ment, many efforts have been exerted to optimize the process of image
acquisition and data processing. One type of efforts mainly focuses on
hardware improvement. In 2009, Grant [15] measured the Young’s mod-
ulus and coefficient of thermal expansion of a nickel-base alloy from
ambient to 1000 °C through combining filters and blue illumination.
The accuracy of strain measurement was enhanced by improving the
quality of speckle pattern through using the natural contrast from the
sample. In 2011, Pan [16] employed the band-pass filter to eliminate the
influence of black-body radiation of high-temperature objects on the in-
tensity of captured images and obtained high-quality digital images at
high temperatures up to 1200 °C. In 2011, Novak [17] obtained accu-
rate measurement of thermal expansion coefficient at temperatures up
to 1500 °C by using an air knife to mitigate the effect of heat haze. In
2016, Chen [18] proposed a two-dimensional ultraviolet digital image
correlation system to obtain high quality speckle pattern by minimizing
the radiation effect and a method using an air controller in combination
with image average algorithm to reduce the effect of heat disturbance at
different temperatures up to 1200 °C. However, the higher the tempera-
ture is, the faster the background radiation increases and the more seri-
ous the airflow disturbance will be. It is difficult to completely suppress
the influence of background radiation and airflow disturbance by only
improving the hardware. It has to seek other solutions to handle with
the problem of poor correlation between images before and after defor-
mation, which will result in random noises in displacement field and
low accuracy of strain measurement. Thus, another type of efforts has
been carried out, aiming at the improvement of the algorithm of strain
measurement. In 2015, Su [19] proposed an effective grayscale-average
technique to minimize the thermal disturbance effect on strain measure-
ment accuracy under high-temperature atmosphere environment. How-
ever, the method of image gray averaging is only suitable for a static
thermal process, such as thermal expansion experiment, in which the
images collected at the same temperature are considered to have no de-
formation. In 2010, Pan [20] proposed a solution for local de-correlated
regions caused by black-body radiation through combining transient
aerodynamic heating simulation device with the reliability-guided digi-
tal image correlation and then measured the thermal deformation field
of the chromium nickel austenite stainless steel sample from room tem-
perature to 550 °C. In 2013, Pan [21] combined a novel active imaging
optical system and a robust reliability-guided displacement tracking al-
gorithm with an automatic reference image updating scheme to extract
full-field thermal deformations of a chromium–nickel austenite stainless
steel sample which was heated from room temperature to 1300 °C and
a woven C/SiC composite at 1550 °C. Although the quality of the image
and the correlation between images before and after deformation can
be improved, the noises in the displacement field can’t be completely
eliminated. Therefore, in order to improve the accuracy of strain mea-
surement, the noisy displacement field should be smoothed firstly.
These existing noisy displacement field smoothing methods can
be divided into two categories, local smoothing methods and global
smoothing methods. The typical local smoothing methods include the
least squares method and its improved versions [22–24]. In the local
smoothing method, only a small part of the displacement data defined
in a calculation window are used for computation by a given fitting func-
tion. The widely used global smoothing methods are the thin-plat spline
smoothing method [25], the finite element based smoothing method
[26] and its improved versions [27,28]. In the global smoothing meth-
ods, all the displacement data points are considered and fitted as a
whole. In both of the two kinds of smoothing methods, the noisy dis-
placement field is smoothed by fitting. When isolated outliers exist, and
especially when abnormal regions which contain many outliers exist,
the performance of fitting will be greatly decreased, because the condi-
tional fitting methods regard the abnormal regions as common data and
the outliers cannot be effectively removed or suppressed.
9
Optics and Lasers in Engineering 111 (2018) 8–18
The extensively used outliers removing method is the “3-𝜎” princi-
ple. However, the “3-𝜎” principle is effective only for isolated outliers
and ineffective for the abnormal regions. Therefore, in order to smooth
the noisy displacement field with isolated outliers and abnormal regions,
new methods should be taken into consideration. Random Sample Con-
sensus algorithm (RANSAC) [29–31] is a powerful fitting tool. Unlike
other smoothing methods by finding a best fitting among all the data
points, RANSAC algorithm removes the outliers and abnormal regions
before smoothing. The classical RANSAC algorithm is a resampling tech-
nique that generates the optimal minimal data model by using the mini-
mum number of data points [32] randomly selected in an iterative way.
If the whole dataset has a large number of points, the instability of the
optimal minimal data model will result in poor robustness of the algo-
rithm. This is the case that has to be faced when the classical RANSAC
algorithm is used for smoothing the noisy displacement field. In this
paper, an Improved Random Sample Consensus (IRANSAC) algorithm
is presented. An adaptive noise threshold selection method for uniform
deformation is put forward. Also, in order to improve the algorithm ro-
bustness, a subset with all the data points after removing the outliers
and abnormal regions, instead of the minimal data model which con-
tains the minimum number of data points, is used to fit the displacement
field plane.
An experiment on tensile test of a C/C composite sample at 2000 °C in
the inert atmosphere was carried out. Laser speckle was used as the char-
acteristic pattern. The DIC method was applied to obtain the displace-
ment field and the IRANSAC algorithm was implemented to remove the
noises and smooth the displacement field before strain calculation. Ex-
perimental results showed that the strain measured by the proposed
method was basically consistent with the result simultaneously mea-
sured by a high temperature contact extensometer. Comparative study
with the method by DIC and least squared algorithm was also carried
out, and the effectiveness of the proposed method was verified.
The rest of this paper is structured as follows. In Section 2, the princi-
ple of DIC method is briefly introduced and the proposed IR-DIC method
is described with a numerically simulated example. In Section 3, an
experiment on room temperature tensile test of stainless steel speci-
men is introduced. In Section 4, an experiment on high temperature
tensile test of a C/C composite sample at 2000 °C in the inert atmo-
sphere is introduced. Finally, the discussions and conclusions are given
in Sections 5 and 6.
2. Basic principles
2.1. Laser speckle
When the laser irradiates the surface of an object with optical rough
surface (compared with the wavelength of light wave), the surface of
the object scatters numerous coherent wavelets, which interfere with
each other in the space around the object. If the phase difference of
the scattered wavelet satisfies the condition of constructive interference,
these scattering wavelets will form a bright spot. On the contrary, if the
phase difference satisfies the condition of cancellation interference, the
dark spot will be formed. Because the phase of the scattered wavelet is
random distribution, a number of bright spots and dark spots formed by
the scattered wavelets are randomly distributed around the surface of
the object. The bright spots and dark spots are called laser speckle.
2.2. Displacement field measurement by DIC
Digital image correlation uses random intensity distributions on the
specimen surface to obtain the full-field displacement by matching sub-
sets of interest before and after deformation. Thus, mechanical prop-
erties of the material can be analyzed. The basic principle of DIC is
schematically illustrated in Fig. 1. The recorded image before defor-
mation is regarded as “reference image”, and the recorded image after
deformation is regarded as “deformed image”. In the reference image,
J. Song et al. Optics and Lasers in Engineering 111 (2018) 8–18

Fig. 1. Schematic principle of digital image correlation.

a subset of (2M + 1) × (2M + 1) pixels centered at a point is chosen as Fig. 2(b). The generation process is as follows: 21 noise data points ran-
the reference subset. These differences between the central position of domly generated within the range from −0.01 to 0.01 pixel are randomly
the reference subset and the target subset in the deformed image are added to a given matrix with the size of 14 × 15; then a matrix of 6 × 5
displacement component u in the x direction and displacement compo- with uniformly distributed data randomly generated within the range
nent v in the y direction. Then, the Newton–Raphson (N–R) method with from 0 to 0.01 pixel is added to the given matrix, which is used to simu-
the bicubic spline interpolation method can be used to obtain the sub- late the abnormal region; the given matrix is interpolated by the bicubic
pixel displacement. The U displacement field composed of displacement spline interpolation method with the coordinate interval of 10 to gen-
components in the x direction and the V displacement field composed erate a noise matrix of 131 × 141. Then the noise matrix of 131 × 141 is
of displacement components in the y direction can be obtained. added to the simulated displacement field to generate a noisy displace-
The zero-mean normalized cross-correlation (ZNCC) criterion is se- ment field as Fig. 2(c). In Fig. 2(d), it can be seen that the displacement
lected as the evaluation function of the similarity between the reference field fitted by the least square algorithm has a deviation from the sim-
image subset and the target image subset in this paper. ulated displacement field because of the effect of noises. And the strain
along the y direction is 218.9 u𝜀. Compared with the theoretical value,
𝐶𝑍𝑁𝐶𝐶 the error of strain value is 18.9 u𝜀. The difference between the displace-
∑𝑀 (
∑𝑀 ) ( ( ) )
𝑥=−𝑀 𝑓 (𝑥, 𝑦) − 𝑓̄ × 𝑔 𝑥′ , 𝑦′ − 𝑔̄
𝑦=−𝑀 ment field fitted by the LS algorithm and the simulated displacement
= √ √ field is shown in Fig. 2(e).
∑𝑀 ∑𝑀 ( )
̄ 2 × ∑𝑀 ∑𝑀
𝑦=−𝑀 𝑓 (𝑥, 𝑦) − 𝑓 𝑦=−𝑀 (𝑔 (𝑥 , 𝑦 ) − 𝑔̄ )
′ ′ 2
𝑥=−𝑀 𝑥=−𝑀
(1) 2.4. Determining strain by random sample consensus(RANSAC)

In Eq. (1), f(x, y) is the gray value at coordinates (x, y) of the ref-
erence image, g(x′, y′) is the gray intensity value at coordinates (x′,
∑ ∑𝑀
y′) of the deformed image, 𝑓̄ = 1∕(2𝑀 + 1)2 𝑀 𝑥=−𝑀 𝑦=−𝑀 𝑓 (𝑥, 𝑦) and
∑𝑀 ∑ 𝑀
𝑔̄ = 1∕(2𝑀 + 1) 2
𝑥=−𝑀 𝑦=−𝑀 𝑔(𝑥 , 𝑦 ) are the mean intensity values of
′ ′
reference and target subsets.
2.3. Determining strain by least square algorithm (LS)
When the displacement field is obtained, the strain value can be ob-
tained by the derivation of the displacement field. For uniform deforma-
tion, the full field least square algorithm is widely used for strain calcula-
tion. The principle of least square algorithm to fit a plane is to determine
the position of the plane by “minimum residual sum of squares”, in or-
der to find the best matching function of data points. For the discrete
displacement field data v(xi , yj ):
( )
𝑣 𝑥𝑖 , 𝑦𝑗 = 𝑏0 + 𝑏1 𝑥𝑖 + 𝑏2 𝑦𝑗 (2)
The residual sum of squares can be calculated as follows:
𝑚 ∑
∑ 𝑛
( )2
𝑆= 𝑏0 + 𝑏1 𝑥𝑖 + 𝑏2 𝑦𝑗 − 𝑣𝑖𝑗 (3)
𝑖=1 𝑗=1
In Eq. (3), b1 , b2 , b3 are coefficients of the fitting plane, vij is the dis-
placement of point (xi , yj ) along y-axis. Thus, the strain component ɛy
in the y direction equals to the coefficient b2 . The calculation of the
displacement field 𝑢(𝑥𝑖 , 𝑦𝑗 ) = 𝑎0 + 𝑎1 𝑥𝑖 + 𝑎2 𝑦𝑗 in the x direction can be
carried out in the same way. The strain component ɛx in the x direction
equals to the coefficient a1 .
In the least square algorithm, all the data points are involved. Noises
in data points will degrade the fitting result. In order to explain the prin-
ciple of the LS algorithm, a simulated displacement field with 131 × 141
pixels is shown in Fig. 2(a). The strain along the y direction is set as
200 u𝜀. Then a lot of random noise data including both isolated out-
liers and abnormal region containing many outliers are generated as
Being different from the least square algorithm which uses as much
of data points as possible to obtain fitting plane, the RANSAC obtains
the fitting plane by removing noises first. The process of the RANSAC is
stated as follows:
(1) Randomly select a minimum of S data points to determine a min-
imal data model Mk , and use the minimal data model Mk to fit
the displacement field plane.
(2) Calculate the distance d from each data point to the model Mk .
(3) Take the data points which distances are less than the threshold
t as the effective data, and all the effective data points make up
a subset pk .
(4) Repeat steps (1) to (3) K times. The minimal data model Mk cor-
responding to the maximum number of effective data points is
recorded as the best minimal data model. The strain can be ob-
tained from the selected best minimal data model.
As shown in Fig. 3(a), there is the subset pk with all the remaining
effective data points screened by RANSAC algorithm. It can be seen that
a lot of noises are removed and the minimal data model Mk consists of
S data points. The smoothed displacement filed determined by the min-
imal data model Mk is shown in Fig. 3(b). The smoothed displacement
field is smoother and more accurate than the displacement filed fitted
by the LS algorithm. As shown in Fig. 3(c), the difference between the
displacement field determined by the RANSAC algorithm and the simu-
lated displacement field is smaller compared with the result in Fig. 2(e).
The strain value calculated by using the smoothed displacement field is
200.1 u𝜀. Compared with the theoretical value, the error of strain value
is 0.1 u𝜀. Therefore, the accuracy of the strain value calculated by using
the smoothed displacement field will be improved.
For the displacement field, there are two difficulties in the classical
RANSAC algorithm. First, the displacement field contains a large amount
of data points, which means that the S data points selected each time
are different. Thus, since the initial model is determined by randomly

10
J. Song et al.
Fig. 2. Schematic principle of LS: (a) The simulated displacement field (b) Imposed random noise data (c) The noisy displacement field (d) The displacement field
fitted by the LS algorithm (e) The difference between the displacement field fitted by the LS algorithm and the simulated displacement field.
selected data points, the selected optimal data models will be different
every time the RANSAC algorithm is implemented. Then if the selected
best minimal data model Mk is used to fit the displacement field, the
robustness of the algorithm can’t be guaranteed. Second, a proper noise
threshold is very important. If the given threshold is small, many effec-
tive data points will be removed. On the contrary, if the given threshold
is large, many outliers and abnormal regions can’t be removed. Both of
the two cases will lead to poor accuracy and robustness of the algorithm.
2.5. The Improved Random Sample Consensus algorithm (IRANSAC)
In order to improve the robustness of the algorithm and adaptively
select the noise threshold in the process of calculation, the classical
RANSAC algorithm is improved in two aspects:
(1) An adaptive noise threshold selection method is proposed. It is
known that in the V displacement field, displacement values in
the same y coordinate (as x coordinate changes) are the same in
theory. Hence, an adaptive noise threshold selection method is
11
Optics and Lasers in Engineering 111 (2018) 8–18
proposed as follows. For V displacement field, the value Z(xi , yj )
calculated by Eq. (4) is defined as the displacement fluctuation
value. These displacement fluctuation values of all data points
are shown in Fig. 4. It can be seen these displacement fluctua-
tion values and those noise values in Fig. 2(b) are very similar.
Therefore, these displacement fluctuation values are taken as an
indicator of evaluating the displacement field noises. As shown
in Fig. 5, the histogram indicates the frequency of each displace-
ment fluctuation value. The displacement fluctuation value cor-
responding to the average frequency of the histogram is taken as
the noise threshold t. The same noise threshold selection method
can be carried out for the U displacement field.
𝑚
( ) ( ) 1 ∑ ( )
𝑍 𝑥𝑖 , 𝑦𝑗 = 𝑣 𝑥𝑖 , 𝑦𝑗 − 𝑣 𝑥𝑖 , 𝑦𝑗 , 𝑖 = 1 ⋯ 𝑚, 𝑗 = 1 ⋯ 𝑛 (4)
𝑚 𝑖=1
(2) Use the subset pk rather than the minimal data model Mk to fit
the displacement field plane. As shown in Fig. 6(a) and (b), two
subsets pk are obtained by different minimal data models Mk . It
can be seen that data points contained in the two subsets are
J. Song et al. Optics and Lasers in Engineering 111 (2018) 8–18

Fig. 3. Schematic principle of RANSAC: (a) The subset pk with remaining effective data points (b) The smoothed displacement filed determined by RANSAC (c) The
difference between the displacement field determined by the RANSAC algorithm and the simulated displacement field.

Fig. 4. Displacement fluctuation values.

basically similar. Using the subset pk to fit the displacement field Fig. 5. Histogram of displacement fluctuation values.
plane can improve the robustness of the algorithm.

In order to verify the validity of the adaptive noise threshold selec-
tion method, the strain of the noisy displacement field in Fig. 2(c) is cal-
culated 10 times for each threshold. Then the mean values and standard
deviation values of strain error under different thresholds are obtained.
As shown in Fig. 7, when the threshold is smaller or larger, mean val-
ues of strain error will increase. And when the threshold is larger, stan-
dard deviation values will also increase. The threshold calculated by the
adaptive noise threshold selection method is 0.0029, which is within the
range of the preferred threshold. Strain error values calculated by dif-
ferent methods for 10 times are shown in Fig. 8. It can be seen from the
box diagram, the strain measurement accuracy obtained by IRANSAC
algorithm is higher than that obtained by LS algorithm. Also, compared
with the RANSAC algorithm, the robustness of the IRANSAC algorithm
is improved.
2.6. The algorithm flow of IR-DIC
The flowchart of the IR-DIC method is presented in Fig. 9. The brief
process of the IR-DIC method and parameters selection are stated as
follows:
Step 1: Prepare laser speckle. The specimen surface is irradiated by
the laser beam after beam expansion by the spatial filter, and
the specimen surface will scatter numerous coherent wavelets.

12
J. Song et al. Optics and Lasers in Engineering 111 (2018) 8–18

Fig. 6. Two subsets pk obtained by different minimum data models Mk .

Fig. 7. Effect of thresholds on strain measurement.

Fig. 8. Comparison of results obtained by different methods.

Then these scattering wavelets interfere with each other in the

surrounding space and laser speckle will be produced.
Step 2: Obtain the noisy displacement field. Use a camera to get
laser speckle patterns before and after deformation and the noisy
displacement field is calculated by the DIC method.
Step 3: Set the maximum number of iteration K. The maxi- Fig. 9. Flowchart of the IR-DIC method.
mum number of iteration can roughly be obtained by 𝐾 =
log(1 − 𝑝)∕log(1 − (1 − 𝑒)𝑆 ), where S is the minimum number of
data points needed to estimate the model, e is the percentage of
outliers in the data points and p is the requested probability of Step 4: Set the minimum number of S. For the 𝐾 =
success. log(1 − 𝑝)∕log(1 − (1 − 𝑒)𝑆 ), the smaller the value S is, the
less the number of iteration needed will be. Because at least
Calculate the threshold t by the adaptive noise threshold selection three points are needed to fit a plane, so the minimum number
method. of S is set as three to build the model.
Set the initial value of n as zero. Step 5: Calculate the distance d of each point to the fitting plane. The
The number of iteration is recorded as k. And set the initial value of number of data points which distances are less than the threshold
k as zero. t is recorded as Nk and the number of iteration is recorded as k.

13
J. Song et al.
Fig. 10. Schematic diagram of the experimental setup.
Fig. 11. Experimental setup.
If the number of data points Nk > n, set 𝑛 = 𝑁𝑘 . Otherwise, set
𝑛 = 𝑛. The iteration is terminated when the iteration k > K.
Step 6: The subset pk corresponding to the maximum value n is taken
as effective data points for fitting the displacement field plane.
Then the strain value is obtained by using the smoothed displace-
ment field.
3. Tensile experiment of 304 stainless steel sample at room
temperature
In order to prove that laser speckle can be applied to DIC, a tensile
experiment for stainless steel specimen at room temperature was car-
ried out. The schematic diagram of the experimental setup is shown in
Fig. 10. And the actual experimental setup is shown in Fig. 11. A single
frequency laser (wavelength: 532 nm, power: 1200 mW), a spatial filter,
a CCD (resolution: 1392 × 1040 pixels) and a computer are included. The
tensile test with a tensile rate of 0.5 mm/min and a tensile time of 50 s
was carried out twice. During the tensile test, both the contact exten-
someter and non-contact optical measuring device by laser speckle were
used to measure the deformation of the specimen simultaneously.
The sample of stainless steel used in the experiment is shown in
Fig. 12. The speckle pattern with size of 200 × 400 pixels is shown in
Fig. 13.
The time-strain curves of each test are calculated respectively. For
the first test, as shown in Fig. 14(a), the time-strain curves measured by
the extensometer and DIC are in good agreement. And the mean error
of strain is 2.6 × 10–5 𝜀. Also, for the second test, as shown in Fig. 14(b),
the time-strain curves measured by extensometer and DIC are basically
identical. And the mean error of strain is 4.1 × 10–7 𝜀. The experimental
results showed that laser speckle can be successfully applied to digital
image correlation for strain measurement.
14
Optics and Lasers in Engineering 111 (2018) 8–18
Fig. 12. stainless steel sample.
Fig. 13. Speckle pattern at room temperature.
4. Tensile experiment of C/C composite sample at 2000 °C
4.1. Experimental procedure and experimental images
A non-contact optical measuring device is designed and the
schematic diagram of the experimental setup is shown in Fig. 15. It
mainly includes a single frequency laser (wavelength: 532 nm, power:
1200 mW), a 532 nm narrow band filter, a spatial filter, a CCD (res-
olution: 1392 × 1040 pixels) and a computer. The non-contact optical
measuring device was set up as shown in Fig. 16 before the experiment.
After the experimental setup was set up, the specimen was heated to
2000 °C by electric heating in the inert atmosphere heating furnace.
And then the tensile test with a tensile rate of 2 mm/min and a tensile
time of 30 s was carried out after heat preservation for 5 minutes. The
deformation of the specimen was controlled in the elastic stage of the
material. In the tensile test, the contact extensometer and non-contact
optical measuring device were used to measure the deformation of the
specimen simultaneously.
The sample of C/C composite used in the experiment is shown in
Fig. 17. A high temperature speckle pattern collected at 2000 °C is shown
in Fig. 18.
4.2. Experimental results
The noisy displacement field obtained from the last speckle pattern
(speckle pattern collected at 30 seconds) is selected to verify the validity
of the IRANSAC algorithm. As shown in Fig. 19(a), the V displacement
field contains a large number of random noises that include both isolated
outliers and abnormal regions. Because the deformation of specimen is
in the elastic stage of the material, the V displacement field varies uni-
formly along y direction. The displacement field fitted by LS algorithm is
shown in Fig. 19(b). The deviation of the displacement field is very large
due to the influence of noises. The IRANSAC algorithm can remove the
J. Song et al. Optics and Lasers in Engineering 111 (2018) 8–18

Fig. 14. Comparison of time-strain curves: (a) the result of the first experiment (b) the result of the second experiment.

Fig. 15. Schematic diagram of the experimental setup.

Fig. 17. C/C composite sample.

Fig. 16. Experimental setup.

noise points and smooth the displacement field by using these remaining
effective data points. In the calculation process, the number of iteration
K is set as 40, and the threshold is selected by adaptive noise threshold
selection method. As shown in Fig. 19(c), the denoised displacement
field is the subset pk with these remaining effective data points that cal-
culated by IRANSAC algorithm. In the subset pk , isolated outliers and
abnormal regions have been removed. And the smoothed displacement
field calculated by IRANSAC is shown in Fig. 19(d). It can be seen that
the smoothed displacement field calculated by the IRANSAC algorithm Fig. 18. Speckle pattern at 2000 °C.
is more consistent with the characteristic of uniform deformation in the
elastic stage of the material than the displacement field calculated by
the LS algorithm. by different methods. It can be seen from the box diagram, the strain
As shown in Fig. 20, the noise threshold t selected by the adap- measurement accuracy obtained by IRANSAC algorithm is higher than
tive noise threshold selection method is 0.2062. The noisy displacement those obtained by the LS algorithm and the classical RANSAC algorithm.
field is calculated 10 times by IRANSAC using different noise thresholds. The robustness of the IRANSAC algorithm is better than that of the clas-
Mean values and standard deviation values of strain error with different sical RANSAC algorithm.
noise thresholds are shown in Fig. 21. It can be seen from Fig. 21, the As shown in Fig. 23, the strain curve calculated by the traditional DIC
threshold selected by the adaptive noise threshold selection method is method based on least square algorithm is more deviant from the strain
within the range of the preferred threshold. curve measured by the contact extensometer. And the strain curve ob-
Strain error values calculated by different methods are shown in tained by the IR-DIC method is basically consistent with the result mea-
Fig. 22. The strain of the noisy displacement field is calculated 10 times sured by the contact extensometer. It can be seen that these strain curves

15
J. Song et al. Optics and Lasers in Engineering 111 (2018) 8–18

Fig. 19. Comparison of LS algorithm and IRANSAC algorithm: (a) The V displacement field (b) The displacement field fitted by LS (c) The subset pk with remaining
effective data points (d) The smoothed displacement field calculated by IRANSAC.

Fig. 21. Effect of threshold on strain measurement.

Fig. 20. Histogram of displacement fluctuation values.

13.94%, which is calculated by Eq. (5).

measured by three methods are basically similar within the first 10 s.
But with the increase of deformation, the deviation between the strain
1 ∑25 || 𝜀𝐼𝑅 − 𝜀𝐿𝑆 ||
𝑖 𝑖
curve calculated by the LS algorithm and the strain curve measured by 𝜀𝑑𝑒 = × | | × 100% (5)
17 𝑖 =9 | 𝑖
𝜀𝑒𝑥 |
the contact extensometer increases, and the strain curve calculated by | |
the IR-DIC method through removing noises from the noisy displace-
ment field is less deviant from the strain curve measured by the contact In Eq. (5), ɛde is the average deviation of 17 data points from 10 to 30 s.
extensometer. Compared with the least square algorithm, the mean de- 𝜀𝑖𝐼𝑅 is the strain value of each point calculated by the IR-DIC method.
viation between the strain curve calculated by the IR-DIC method and 𝜀𝑖𝐿𝑆 is the strain value of each point calculated by the LS algorithm. 𝜀𝑖𝑒𝑥
the strain curve measured by the contact extensometer is reduced about is the strain value of each point measured by the contact extensometer.

16
J. Song et al.
Fig. 22. Comparison of results obtained by different methods.
Fig. 23. Comparison of strain curves obtained by different methods.
5. Discussions
An adaptive noise threshold selection method for uniform deforma-
tion is put forward to update the threshold during the process of smooth-
ing the displacement field to solve the problem of threshold selection.
As shown in Figs. 7 and 21, the threshold calculated by the adaptive
noise threshold selection method is within the range of the preferred
threshold. Although the threshold obtained by the adaptive threshold
selection method can’t guarantee to be the optimal threshold, it is close
to the optimal value and can effectively improve the accuracy of strain
Fig. 24. The thresholds corresponding to different number of groups: (a) Results of the simulated displacement field (b) Results of the displacement field of high
temperature experiment.
17
Optics and Lasers in Engineering 111 (2018) 8–18
measurement. It needs further research in the future to directly find the
best threshold value to obtain the higher accuracy strain measurement.
When calculating the histogram, groups of displacement fluctuation
values should be divided first. The thresholds corresponding to different
number of groups are shown in Fig. 24(a) and (b). The results of the
simulated displacement field are shown in Fig. 24(a). And the results
of the displacement field of high temperature experiment are shown in
Fig. 24(b). It can be seen that the size of the threshold is insensitive to the
number of groups. And the thresholds calculated by different number of
groups are all in the optimal threshold selection range. In the calculation
process, the number of groups is set as 100 for all displacement fields in
the experiments.
The IRANSAC algorithm can only be applied to uniform deforma-
tion, because the adaptive noise threshold selection is only suitable for
uniform deformation. And the stability of the IRANSAC algorithm has
been greatly improved. As shown in Figs. 8 and 22, when strain values
calculated by the IRANSAC algorithm and RANSAC algorithm for 10
times respectively, the fluctuation of strain error values calculated by
the IRANSAC algorithm is smaller than that calculated by the RANSAC
algorithm.
In practice, the probability of success and the percentage of outliers
in Step 3 of the proposed method are unknown. The general settings are
p = 99% as the probability of success and e = 50% as the percentage of
outliers. It can be seen from the formula 𝐾 = log(1 − 𝑝)∕log(1 − (1 − 𝑒)𝑆 ),
the probability of success is proportional to the number of iteration K.
So, if data processing capacity is strong enough, the number of iteration
should be increased to improve the accuracy of strain measurement.
The problem of de-correlation should be considered when laser
speckle is used in digital image correlation for large deformation mea-
surement. The de-correlation of laser speckle relates to many factors,
such as experimental temperature and specimen material. In the ten-
sile experiment at room temperature, when the strain is greater than
2000 u𝜀, the laser speckle pattern de-correlation can be observed and
strain measurement accuracy begins to decrease. In the tensile experi-
ment at high temperature, when the strain is greater than 350 u𝜀, the
laser speckle pattern de-correlation can be observed and strain measure-
ment accuracy begins to decrease. The technique of updating reference
image frames [21] which has been proved effective for artificial speckle,
is also applicable to laser speckle and was adopted in the analysis of the
experimental data. However, if reference frames are updated frequently,
the accumulated errors will occur. So, how to solve the de-correlation
problem when using laser speckle for large deformation measurement
is still an ongoing issue.
Strain measurement can be divided into contact measurement
method and non-contact optical measurement method. Although con-
J. Song et al.
tact strain measurement method has a stable measurement result, there
are some difficulties in using contact measurement method for high
temperature strain measurement, such as the selection of the special
heat-resisting material for extensometer, the vulnerability characteris-
tics and the error compensation for thermal deformation of the exten-
someter. The DIC method, as one of the non-contact optical measure-
ment method, if the noise problem that caused in the high temperature
environment can be effectively addressed, will be an excellent alterna-
tive to the contact strain measurement.
6. Conclusions
In ultra-high-temperature environment, the problem of speckle
preparation is a serious problem that the non-contact deformation mea-
surement has to face. A speckle pattern preparation method by laser
speckle is proposed for deformation measurement at high temperature.
With laser speckle pattern, it is not necessary to seek special adhesive or
speckle paint resistant to high temperature. A high temperature strain
measurement method (IR-DIC) by combining digital image correlation
of laser speckle and Improved Random Sample Consensus Smoothing
algorithm (IRANSAC) for uniform deformation is proposed to smooth
the noisy displacement field. First, rather than using the minimal data
model in the classical RANSAC algorithm, a subset with all the remain-
ing effective data points after removing outliers and abnormal regions
is used to fit the displacement field plane. Second, an adaptive noise
threshold selection method for uniform deformation is proposed to up-
date the threshold to adapt to different noisy displacement fields.
Finally, in order to verify the effectiveness of the proposed method,
a high temperature tensile test of C/C composite at 2000 °C in the inert
atmosphere had been carried out. The strain curve calculated by the
IR-DIC method is basically consistent with the result measured by the
contact extensometer, which is better than the strain curve calculated
by DIC with the LS algorithm.
Acknowledgments
This work was supported by the Special-funded Program on Na-
tional Key Scientific Instruments and Equipment Development (grant
no. 2011YQ14014507). The authors would also like to thank the edi-
tors and reviewers for their valuable suggestions and kind work.
References
[1] Wu R, Kong C, Li K, Zhang D. Real-time digital image correlation for dynamic strain
measurement. Exp Mech 2016;56:833–43.
[2] Sutton MA, Hild F. Recent advances and perspectives in digital image correlation.
Exp Mech 2015;55:1–8.
[3] Lyons JS, Liu J, Sutton MA. High-temperature deformation measurements using dig-
ital-image correlation. Exp Mech 1996:64–70.
[4] Liu ZW, Gao JX. Deformation-pattern-based digital speckle correlation for coefficient
of thermal expansion evaluation of film. Opt Exp 2011;19:17469–79.
[5] Dong J, Liu Z, Gao J. Multi-parameter inversion and thermo-mechanical deformation
decoupling using I-DIC. Exp Mech 2017;57(1):31–9.
18
Optics and Lasers in Engineering 111 (2018) 8–18
[6] Fang XF, Jia JM, Feng X. Three-point bending test at extremely high temperature
enhanced by real-time observation and measurement. Measurement 2015;59:171–6.
[7] Chen X, Xu N, Yang LX, Xiang D. High temperature displacement and strain mea-
surement using a monochromatic light illuminated stereo digital image correlation
system. Meas Sci Technol 2012;23:125603.
[8] Yang XB, Liu ZX, Xie HM. A real time deformation evaluation method for surface
and interface of thermal barrier coatings during 1100°C thermal shock. Meas Sci
Technol 2012;23:105604.
[9] Goodman JW. Laser speckle and related phenomena. Springer; 1984.
[10] Yamaguchi I. Speckle displacement and decorrelation in the diffraction and image
fields for small object deformation. Opt Acta 1981;28(10):1359–76.
[11] LePage WS, Shaw JA, Daly SH. Optimum paint sequence for speckle patterns in
digital image correlation. Exp Tech 2017;41:557–63.
[12] Lagattu F, Brillaud J, Lafarie-Frenot MC. Progress in mechanics of materials by using
laser speckle method. Proc IUTAM 2000;82:635–42.
[13] Anwander M, Zagar BG, Weiss B, Weiss H. Noncontacting strain measure-
ments at high temperatures by the digital laser speckle technique. Exp Mech
2000;40(1):98–115.
[14] Brillaud J, Lagattu F. Limits and possibilities of laser speckle and white-light image–
correlation methods: theory and experiments. Appl Opt 2002;41(31):6603–13.
[15] Grant BMB, Stone HJ, Withers PJ, Preuss M. High-temperature strain field measure-
ment using digital image correlation. J Strain Anal 2009;44:263–71.
[16] Pan B, Wu DF, Wang ZY, Xia Y. High-temperature digital image correlation
method for full-field deformation measurement at 1200°C. Meas Sci Technol
2011;22:015701.
[17] Novak MD, Zok FW. High-temperature materials testing with full-field strain mea-
surement: Experimental design and practice. Rev Sci Instrum 2011;82:115101.
[18] Chen L, Wang YH, Dan XZ, Xiao Y, Yang LX. Experimental research of digital image
correlation system in high temperature test. SPIE 2016;9903:990306.
[19] Su YQ, Yao XF, Wang S, Man YJ. Improvement on measurement accuracy of high-
-temperature DIC by grayscale-average technique. Opt Lasers Eng 2015;75:10–16.
[20] Pan B, Wu DF, Xia Y. High-temperature deformation field measurement by combin-
ing transient aerodynamic heating simulation system and reliability-guided digital
image correlation. Opt Lasers Eng 2010;48:841–8.
[21] Pan B, Wu DF, Gao JX. High-temperature strain measurement using active
imaging digital image correlation and infrared radiation heating. J Strain Anal
2013;0(0):1–9.
[22] Pan B, Qian KM, Xie HM, Asundi A. Two-dimensional digital imagecorrelation
for in-plane displacement and strain measurement: a review. Meas Sci Technol
2009;20(6):062001.
[23] Pan B, Asundi A, Xie HM, Gao JX. Digital image correlation using iterative least
squares and pointwise least squares for displacement field and strain field easure-
ments. Opt Lasers Eng 2009;47:865–74.
[24] Dai XG, Yang FJ, Chen ZN, Shao XX, He XY. Strain field estimation based on digital
image correlation and radial basis function. Opt Lasers Eng 2015;65:64–72.
[25] Wang CC, Deng J, Ateshian GA, Hung CT. An Automated approach for direct mea-
surement of two-dimensional strain distributions within articular cartilage under
unconfined compression. J Biomech Eng 2002;124(5):557–67.
[26] Meng LB, Jin GC, Yao XF. Application of iteration and finite element smoothing
technique for displacement and strain measurement of digital speckle correlation.
Opt Lasers Eng 2007;45(1):57–63.
[27] Zhao J, Zeng P, Pan B. Improved Hermite finite element smoothing method for ful-
l-field strain measurement over arbitrary region of interest in digital image correla-
tion. Opt Lasers Eng 2012;50:1662–71.
[28] Yoneyama S. Smoothing measured displacement and computing strains utilising fi-
nite element method. Strain 2011;47(2):258–66.
[29] Fischler MA, Bolles RC. Random sample paradigm for model fitting with ap-
plications to image analysis and automated cartography. Graph Image Process
1981;24(6):381–95.
[30] Misra I, Moorthi SM, Dhar D, Ramakrishnan R. An automatic satellite image regis-
tration technique based on Harris corner detection and random sample consensus
(RANSAC) outlier rejection model. IEEE; 2012.
[31] Liu KX, Qian J, Yang RK. Block matching algorithm based on RANSAC algorithm.
IEEE; 2010.
[32] Derpanis KG. Overview of the RANSAC algorithm. Technical report. Computer Sci-
ence, York Universit; 2010.