Measurement 173 (2021) 108618

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Assessment of speckle pattern quality in digital image correlation from the

perspective of mean bias error
Xiaoliang Hu , Zhijiang Xie , Fei Liu *
State Key Laboratory of Mechanical Transmissions, Chongqing University, Chongqing 400044, China
College of Mechanical Engineering, Chongqing University, Chongqing 400044, China

A R T I C L E I N F O A B S T R A C T

Keywords: Digital image correlation (DIC) technology is severely affected by the speckle pattern quality. Most of the existing
Digital image correlation assessment methods always use only one parameter to evaluate both mean bias error and standard deviation
Speckle pattern quality error, such as the mean intensity gradient (MIG). However, the principles of these two error models are quite
Mean bias error
different. The mean bias error is closely related to the first-order and second-order gradients of speckle pattern
Assessment parameter
intensity. A parameter named Ef, based on the mean bias error, is proposed to evaluate the speckle pattern quality
in this work. Numerical translations are applied to eight different speckle patterns to justify its correctness. The
results indicate that the existing MIG is efficient to assess speckle pattern quality by evaluating the standard
deviation error, while the parameter Ef is efficient to assess the speckle pattern quality from the perspective of
mean bias error.

1. Introduction measurement methods used in experimental mechanics and based on

digital image correlation, simulated speckle patterns were also discussed
Digital image correlation (DIC) is a non-contact technology for full- in Refs. [11,12]. In addition, speckle patterns used in microscale ap­
field displacement and strain measurement. Combined with computer plications and biology were respectively studied by Dong et al. [13] and
vision, it has been widely used in experimental mechanics [1,2]. In Lionello et al. [14]. Speckle pattern optimization includes speckle size,
essence, many factors can affect the accuracy of measurement, such as density, randomness, and intensity distribution. Chen et al. [15] utilized
the speckle pattern quality [3,4], image noise [5,6], and subset size the root-mean-square error and the average number of iterations to
[7,8]. optimize the speckle patterns that consist of single-radius or multi-
The DIC method relies on the speckle pattern placed on an object’s radius circular speckles with different randomness factors of digital
surface. The randomly distributed speckle pattern is as a carrier of speckle patterns. The 2-pixel single-radius speckle pattern with a
deformation information and deforms together with the specimen sur­ random value of 0.3 was recommended. A statistical model was built by
face. Apparently, the measurement errors are directly related to the Su et al. to optimize speckle patterns that are composed of randomly
speckle pattern quality. Much research has been conducted on speckle positioned speckles [16]. Later, Su et al. [17] analyzed the performance
patterns, including speckle pattern fabrication, speckle pattern optimi­ of non-overlapping circular speckles in terms of uniqueness, accuracy,
zation, and speckle pattern assessment. Chen et al. [9] proposed a precision, and spatial resolution. Bomarito et al. [18] optimized speckle
method based on water transfer printing to transfer speckle patterns to patterns according to a composite quality assessment metric that com­
carriers’ surfaces. In their analysis, the speckle patterns were generated bines the sum of the square of subset intensity gradients (SSSIG) [7] and
and optimized by computer simulations. Zhang et al. [10] used an at­ the secondary auto-correlation peak height.
omization system to generate repeatably a speckle pattern that produces The speckle patterns and even the optimized speckle patterns that
the least DIC error. The optimal speckle pattern was obtained by are made using different techniques or by different people may have
adjusting the paint solution delivery speed, the distance between the distinctly different characteristics, including image contrast, gray-level
inner layer and the tip of the atomization probe, and the argon gas flow histogram distribution, and morphological features. Thus, the assess­
rate. To evaluate the metrological performance of displacement ment of speckle pattern quality is crucial when using the DIC method.

* Corresponding author at: State Key Laboratory of Mechanical Transmissions, Chongqing University, Chongqing 400044, China.
E-mail address: [email protected] (F. Liu).

https://doi.org/10.1016/j.measurement.2020.108618
Received 9 August 2020; Received in revised form 19 September 2020; Accepted 12 October 2020
Available online 20 October 2020
0263-2241/© 2020 Elsevier Ltd. All rights reserved.
X. Hu et al. Measurement 173 (2021) 108618

Many parameters are applied to evaluate the speckle pattern quality subsets before and after deformation by different strategies. The surface
based on the morphology, gray-level intensity, and such factors in of the specimen must have a random distributed speckle pattern and the
combination. Pan et al. [19] proposed the mean intensity gradient speckle pattern deforms together with the specimen surface. A correla­
(MIG), which is the most widely used evaluation parameter, based on tion criterion is predefined to evaluate the similarity between the
the SSSIG. The MIG indicates that speckle patterns with a larger MIG will reference and deformed subsets. Forward additive Gauss–Newton al­
produce smaller measurement errors because the mean bias error and gorithm (FA-GN) and IC-GN algorithm are widely used to calculate the
standard deviation error are inversely related to SSSIG. A similar optimal deformation parameters. Once the correlation criterion satisfies
parameter, called mean subset fluctuation, was also proposed by Hua the defined convergence condition, the displacements of the subset
et al. [20]. Zhao et al. [21] used the mean directional intensity deriva­ central can be determined. The basic principle of the IC-GN algorithm
tive (MID) to qualify wood texture pattern qualities for digital image used in this work is shown in Fig. 1. Details of different DIC algorithms
correlation. The principle of parameter MID is similar to MIG but with are described in Ref. [29]. The zero-mean normalized sum of squared
direction. The MID guided the operator to choose a suitable subset size difference (ZNSSD) criterion is used to evaluate the similarity between
in both x- and y-directions. Furthermore, Yu et al. [22] utilized the mean the reference subset and the deformed subset with size of (2M + 1) ×
intensity of the second derivative (MIOSD) to evaluate the speckle (2M + 1) pixels using the following expression:
pattern quality. The results indicated that high-quality speckle patterns { }2
should have a high-intensity gradient and a small second derivative, ∑ ∑
M M
f (W(x, y; Δp)) − f g(W(x, y; p)) − g
CZNSSD (Δp) = − (1)
while the reason was not theoretically explained. The numerical ex­ y=− M x=− M
Δf Δg
periments compared the speckle patterns with the same MIG or the
similar variation tendency between MIG and MIOSD, while the qualities Where f and g are the gray intensities of the reference subset and the
of speckle patterns with different MIG or different variation tendency deformed subset, respectively. p is the pre-computed first-order defor­
between MIG and MIOSD were not known. Liu et al. [23] used Shannon mation parameter vector exerted on the deformed subset. W(x, y; p) and
entropy (SE), which is based on information theory to assess the speckle W(x, y; Δp) are the warp function and incremental warp function of the
pattern quality. The simulations indicated that a good quality speckle deformed subset and reference subset, respectively. f and g denote the
pattern with larger Shannon entropy will produce smaller mean bias mean intensity values of the reference subset and the deformed subset,
error. Two parameters related to the root mean square error (RMSE) √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑y=M ∑x=M ( )2̅
were proposed to characterize the total system and random errors [24]. respectively. Δf = y=− M x=− M f(W(x, y; Δp)) − f , Δg =
The parameters are related to the external noises of reference and √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑y=M ∑x=M 2
deformed images because the error RMSE is related to the noise and y=− M x=− M ((g(W(x, y; p)) − g) .

subpixel translation. Recently, many composite pattern quality metrics

have been applied to assess the quality of the speckle pattern. Park et al. 3. Assessment parameters
[25] introduced a metric called SDGIS to assess the quality of speckle
patterns. The SDGIS was based on the cumulative distribution of the Many parameters were used to evaluate two aspects of speckle
standard deviation of gray intensities within each speckle. Simulations pattern quality: morphology and gray-level intensity. However, few of
and experiments indicated that the parameter SDGIS is inversely pro­ these parameters elucidated the direct interrelationship between the
portional to the averaged error of the subset method. A multi-factor DIC errors and the quality assessment parameters, especially for the
fusion index (MFFI), which considered the inhomogeneity of gray dis­ assessment parameters based on morphology [4,30]. Several speckle
tribution (IGD), mean square deviation of gray (MSDG), and standard pattern quality metrics proposed in the literatures will be described
deviation of speckle particle sizes (SDSPS), was also proposed to assess herein. A new assessment parameter based on the gray-level intensity
the quality of the laser speckle pattern [26]. Other parameters based on gradient and the second derivative of gray-level intensity is proposed.
the morphological features’ statistical properties were discussed in
[27,28].
As observed from Ref. [6], the DIC errors are substantially related to
the gray intensity of the speckle pattern, and the MIG is used to evaluate
the quality of the speckle pattern from the standard deviation [19]. Shao
et al. [5] indicated that the inverse compositional Gauss–Newton algo­
rithm (IC-GN) algorithm, which was more robust to noise, did not
introduce noise-induced bias if the gray gradient operator was chosen
properly. Therefore, if the IC-GN algorithm is used, the noise term in the
mean bias error formula can be ignored. A metric (Ef) is proposed based
on the intensity gradient and the second derivative to evaluate the
speckle pattern quality from the mean bias error. Theoretical derivation
of the metric is given for both the linear interpolation and the cubic
interpolation. This metric is compared to several common assessment
parameters, as discussed above. To validate the correctness and effec­
tiveness of the proposed parameter for speckle pattern quality assess­
ment, eight different speckle patterns obtained from actual experiments
with different intensity distributions and morphologies are used in the
numerical studies. Rigid-body translations are applied to the experi­
ments. It is shown that the mean bias error of the measured displacement
is closely related to the intensity gradient and the second derivative of
the speckle pattern used. Speckle patterns with small Efs will produce
small mean bias errors.

2. Digital image correlation

The fundamental principle of DIC is to match the corresponding Fig. 1. Standard subset-based DIC method with IC-GN algorithm.

2
X. Hu et al. Measurement 173 (2021) 108618

3.1. Existing parameters where W and H (pixels) are the width and height of image, respectively.
⃒ ( ) ⃒ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
⃒∇f xij ⃒ = ∇fx (xij )2 + ∇fy (xij )2 is the modulus of the intensity
3.1.1. Auto-correlation peak sharpness
Bossuyt [30] introduced the autocorrelation peak sharpness radius to gradient at pixel xij . ∇fx (xij ) and ∇fy (xij ) are the x- and y-directional
quantify the sensitivity and robustness of DIC measurement. The metric intensity gradients at pixel xij , respectively, which can be computed by
is defined as: using a central difference algorithm. From the view of standard devia­
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ tion, MIG is useful to evaluate the quality of speckle pattern because the
Rpeak =
A|0 − A∞
〈 〉 (2) standard deviation error of measured displacement is just related to the
A|0 − A|±1 SSSIG and the standard deviation of noise. However, the MIG cannot
〈 〉 evaluate the speckle pattern quality accurate from the perspective of
where A|0 is the autocorrelation value for zero displacement, A|±1 is mean bias error.
the mean autocorrelation value obtained by shifting the pixel position of
the primary peak pixel by one pixel in four cardinal directions, and A|∞ 3.1.4. Multi-factor fusion index
is the value of autocorrelation to which the resulting radius corresponds. The MFFI [26] synthesizes the effects of the gray distribution of the
In the analysis of this work, A|∞ = 0 [18,30]. The parameter Rpeak in­ speckle pattern, the contrast of the speckle pattern and the morphology
dicates that speckle pattern with higher contrast, more features and of speckle particles on the quality of speckle pattern. These three terms
sharper edges performs better in DIC measurement. of MFFI are defined as IGD, MSDG and SDSPS, respectively.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√ L− 1
3.1.2. Shannon entropy √∑
δIGD = √ P2k /(W × H) (7)
The SE for evaluating the quality of speckle pattern was proposed by k=0
Liu [23], which was a measure of the uncertainty associated with a
random variable in the information theory. The SE of image quantifies √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√ L− 1
the expected value of the image information content. It is defined as δMSDG
√∑
= √ {(k − P) × Qk } (8)
follows: k=0

∑
L− 1
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
H(Y) = − p(k)log(p(k)) (3) √ T
√1 ∑
k=0 δSDSPS =√ (An − A)2 (max(An ) > 1) (9)
T n=1
where H is the Shannon entropy, bits/pixel, L is the number of gray
levels, p(k) is the probability of each gray level of the image. The results where Pk is the number of pixels contained in each gray level, Qk = Pk /
show that the DIC measurement accuracy is closely related to the SE of (W × H), P = L−
∑ 1
k=0 k × Qk , An is the size of every speckle particle, A =
image. Larger SE denotes that the speckle pattern has more feature in­ ∑T
An /T is the mean value of speckle pattern size, and T is the number
formation and gives better measurement accuracy. n− 1
of speckle particles. The MFFI is defined as:
3.1.3. Mean intensity gradient 100 × δIGD
δMFFI = (10)
The MIG proposed by Pan et al. [19] is based on the mathematical δMSDG × δSDSPS
expectation and standard deviation of the measured displacement [6].
It is obvious that the smaller the value of MFFI is, the better the
As derived in Ref. [19], both the mean bias error and standard deviation
quality of speckle pattern is.
error of the measured displacement are in inverse proportion to the
SSSIG [7]. The mean bias error and standard deviation error of one-
dimensional (1D) displacement are described as follows: 3.2. Combined intensity gradient and second derivative
∑x=M ∑y=M [ ] 2 2
y=− M − h(xij )∇fx (xij ) + (1 − 2τx )(2M + 1) σ As shown in Eq. (4), if the IC-GN algorithm is used, the mean bias
(4)
x=− M
E(ue ) ≅ ∑x=M ∑y=M [ ]2
x=− M y=− M ∇fx (xij ) error of the estimated displacement is related to the subpixel displace­
ment, the intensity gradient and the h(xij ). As observed from Ref. [6],
√̅̅̅
2σ h(xij ) in Eq. (4) is induced by intensity interpolation. We will derive the
std(ue ) ≅ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (5)
∑ ∑ [
x=M y=M ]2 approximate h(xij ) using two interpolation methods with translated
∇fx (xij ) images. E(ue ) can be explained in the following two subsections:
x=− M y=− M

3.2.1. Linear interpolation

where h(xij ) = g(ξij ) − f(xij ) is the interpolation difference between
For the 1D linear interpolation, the gray intensity at a non-integer
reference and translated images, ξij is the exact position of xij in the
location is estimated by the intensity values at two neighboring
translated image , τx is the sub-pixel displacement, σ 2 is the variance of
integer locations. If the reference image translates τx in the x-direction,
noise and ∇fx (xij ) is the x-directional intensity derivative at pixel xij . It
0 < τx < 1, then the non-integer positions of the reference and the
should be noted that the ∇gx (xij ) of mean bias error and standard de­
translated images can be calculated via linear interpolation, as shown in
viation error in Ref. [6] is replaced by ∇fx (xij ), as shown in Eqs. (4) and Fig. 2.
(5). The gradient ∇gx (xij ) is estimated at non-integer position in the In Fig. 2, all the integer pixels of the reference image translate τx in
deformed image and gradient ∇fx (xij ) is estimated at integer position in the x-direction. There are two steps to obtain the intensity value of the
the reference image. non-integer pixel position in the translated image:
The SSSIG is an effective parameter for evaluating the speckle
pattern quality of local subset. To evaluate the entire speckle quality of (a) The intensity value of the integer pixel location of the translated
image, the MIG is proposed and defined as: image is obtained by calculating the non-integer pixel location in
W ∑
∑ H ⃒ the reference image via linear interpolation.
( ) ⃒/
δf = ⃒∇f xij ⃒ (W × H) (6) (b) The intensity value of the non-integer pixel location in the
i=1 j=1 translated image is obtained by using the intensity values of the
adjacent two integer pixel locations of the translated image.

3
X. Hu et al. Measurement 173 (2021) 108618

Fig. 3. Principle of the intensity value calculation at the non-integer position in

the deformed image via cubic interpolation.

d = f(i− 1)j , (14d)

Fig. 2. Principle of the intensity value calculation at the non-integer position in
the deformed image via linear interpolation.
gij = a(1 − τx )3 + b(1 − τx )2 + c(1 − τx ) + d (15)
As shown in Fig. 2, the intensity value gij is equal to the intensity g(i− 1)j , g(i+1)j , and g(i+2)j can also be calculated by using the formulas in
value f(i− τx )j , which is estimated by f(i− 1)j and fij using a linear interpo­ Eqs. (14a)–(14d), which are similar. Then, the intensity value g(i+τx )j can
lation method. Similarly, the intensity value g(i+1)j can also be obtained. be obtained via cubic interpolation.
Then, the intensity value of fij in the translated image g(i+τx )j is estimated The simplified intensity difference can be obtained with any τx ,
by gij and g(i+1)j via linear interpolation. 0 < τx < 1:

gij = (fij − f(i− 1)j )(1 − τx ) + f(i− 1)j , (11a) h(xij ) = g(1+τx )j − fij
= l1 (f(i− 1)j − 2fij + f(i+1)j ) + l2 (f(i− 2)j − 2fij + f(i+2)j ) + l3 (f(i− 3)j − 2fij + f(i+3)j )
g(i+1)j = (f(i+1)j − fij )(1 − τx ) + fij , (11b) ≅ k∇2 fx (xij )
(16)
g(i+τx )j = (g(i+1)j − gij )τx + gij , (11c) 2( )
where l1 = τ − τx ) − 15τ + 15τx + 12 ,
1 2
4 x
(1 2
x l2 =
The intensity difference h(xij ) can be obtained. 1 2
τ (1 − τ )2(
6τ 2
− 6 τ − 3
)
and l = 1 2
τ (1 − τ )2(
− τ 2
+ 6
)
are the co­
4 x x x x 3 4 x x x

h(xij ) = g(i+τx )j − fij efficients of different second-order central differences obtained by

[( ]
= f(i+1)j − 2fij + f(i− 1)j )(1 − τx ) τx (12) simplifying h(xij ), and k = l1 + l2 + l3. As observed from Eqs. (12) and
= ∇2 fx (xij )(1 − τx )τx (16), the intensity difference is related to the second-order central
difference.
where ∇2 fx (xij ) is the second-order central difference in the x-direction Taking Eq. (16) into Eq. (4) and ignoring the influence of noise, the
at pixel xij . biased expectation in the x-direction can be reduced to:
Taking Eq. (12) into Eq. (4) and ignoring the influence of noise, the ∑x=M ∑y=M [ 2 ]
biased expectation in the x-direction can be reduced to y=− M ∇ fx (xij )∇fx (xij )
(17)
x=− M
E(ue ) ≅ K2 ∑x=M ∑y=M [ ]2
∑x=M ∑y=M [ 2 ] x=− M y=− M ∇fx (xij )
y=− M ∇ fx (xij )∇fx (xij )
(13)
x=− M
E(ue ) ≅ K1 ∑x=M ∑y=M [ ] 2 Eqs. (13) and (17) explain why speckle patterns with a small mean
y=− M ∇fx (xij )
intensity of the second derivative produced small displacement mea­
x=− M

surement errors when the same MIG was used for the speckle patterns as
3.2.2. Cubic interpolation
indicated in Ref. [22].
For the 1D cubic interpolation, the intensity value at a non-integer
location is estimated by the intensity values at four integer locations.
3.2.3. Assessment parameter
Similar to linear interpolation, the non-integer positions of the reference
As observed in Eqs. (13) and (17), the biased expectations of a certain
and translated images can be calculated via cubic interpolation, as
subset obtained by different interpolation methods have similar for­
shown in Fig. 3.
mulas. It should be noted that the parameters K1 and K2 are related to
In Fig. 3, all the integer pixels of the reference image translate τx in
the translation distance τx which can be regarded as constants for all the
the x-direction. There also are two steps to obtain the intensity value of
compared speckle patterns. According to the bias expectation in the x-
non-integer pixel positions in the translated image. As shown in Fig. 3,
direction, the following two parameters in both the x- and y-directions
the intensity value gij is equal to the intensity value f(i− τx )j , which is
for the entire speckle pattern are estimated as follows:
estimated by f(i− 2)j , f(i− 1)j , fij , and f(i+1)j using a cubic interpolation
∑W ∑H ⃒⃒ 2 ⃒
method. Similarly, the intensity value g(i− 1)j , g(i+1)j , and g(i+2)j also can be j=1 ∇ fx (xij )∇fx (xij )
⃒
(18a)
i=1
kx = ∑W ∑H [ ]2
obtained. Then, the intensity value of fij in the translated image, g(i+τx )j ,
i=1 j=1 ∇fx (xij )
is estimated by g(i− 1)j , gij , g(i+1)j , and g(i+2)j via cubic interpolation. For the
∑W ∑H ⃒⃒ 2 ⃒
intensity gij , j=1 ∇ fy (xij )∇fy (xij )
⃒
(18b)
i=1
ky = ∑W ∑H [ ]2
1 3 3 1 j=1 ∇fy (xij )
a = − f(i− 2)j + f(i− 1)j − fij + f(i+1)j , (14a) i=1
2 2 2 2
here ∇fx (xij ) and ∇fy (xij ) are the x- and y-directional intensity gra­
5 1 dients at pixel xij , respectively, ∇2 fx (xij ) and ∇2 fy (xij ) are the x- and y-
b = f(i− 2)j − f(i− 1)j + 2fij − f(i+1)j , (14b)
2 2 directional second-order central differences respectively at pixel xij .
Based on the parameters kx and ky , a new global parameter, namely,
1 1
c=− f(i− 2)j + fij , (14c) the combined intensity gradient and second-order derivative (denoted
2 2
by Ef), is proposed to evaluate the quality of the entire speckle pattern as

4
X. Hu et al. Measurement 173 (2021) 108618

follows: speckles with a one-pixel radius. Next, the speckle particle size was
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅ calculated for its four connected regions. The binarization threshold of
Ef = k2x + k2y (19) speckle pattern (g) was an approximate value selected artificially due to
the unobvious speckle features. As observed in Fig. 6, the change rules
where the parameter Ef is related to the intensity gradient and the for the curves of the SE, MIG, and MFFI criteria are quite different from
second derivative. The intensity gradient can be calculated by using a speckle pattern (a) to speckle pattern (h). The SE indicates that the order
central difference algorithm and the second derivative can be calculated of speckle patterns, from good quality to poor quality, is a > f > c > b >
by using a second-order central difference algorithm. Eq. (19) denotes e > d > g > h; the MIG indicates that the order of speckle patterns, from
that the smaller the value of Ef, the better the speckle pattern’s quality. good quality to poor quality, is a > f > b > c > e > h > d > g; and the
MFFI indicates that the order of speckle patterns, from good quality to
4. Numerical experiment poor quality, is a > b > c > f > e > d > g > h.
As observed in Figs. 5 and 6, the SE, MIG, and MFFI metrics did not
4.1. Speckle patterns provide the same order of speckle pattern quality. This indicates that
different metrics may get different orders of speckle pattern quality
In the numerical experiments, eight 8-bit (0–255 Gy-level range) because of the different care indexes.
speckle patterns with a resolution of pixels were introduced to study the Fig. 7 shows the proposed index Ef of the eight speckle patterns. As
efficiency of the parameter Ef on the assessment of speckle pattern also observed in this figure, the change rule of the curve of Ef is mono­
quality. The eight speckle patterns were obtained by different technol­ tonically increasing from speckle pattern (a) to speckle pattern (h),
ogies, as shown in Fig. 4. indicating that the order of speckle patterns, from good quality to poor
The speckle pattern in Fig. 4(b) was made using a marker pen, Fig. 4 quality, is a > b > c > d > e > f > g > h.
(g) was made by painting, and the other six speckle patterns were made In Fig. 6, the SE, MIG, and MFFI metrics show that speckle patterns
by airbrush with different colored backgrounds and speckles. The in­ (e) and (f) have better speckle quality than speckle pattern (d), while
tensity was determined by the ratio of black and white paint, and the according to Ef in Fig. 7, the quality of speckle pattern (d) is better. As
speckle size and density were respectively determined by airbrush and shown in Fig. 4, speckle patterns (e) and (f) have more morphological
spraying time. As observed in Fig. 4, the gray-level intensity distribu­ features and shape edges than speckle pattern (d), so the SE which based
tions and morphological features of the eight speckle patterns are on feature information and the MIG which based on gradient are larger
distinctly different. than that of speckle pattern (d). More feature information and larger
gradient also will lead to smaller MFFI, and that makes speckle patterns
4.2. Metrics analysis (e) and (f) have better speckle quality than speckle pattern (d). While
more morphological features and shape edges may produce larger
As discussed in Section 3, the parameter Rpeak can evaluate the image interpolation errors and lead to larger Ef , as shown in Fig. 7.
contrast and morphological feature of pattern. The autocorrelation and
the autocorrelation peak sharpness radii of the eight speckle patterns are
shown in Fig. 5. Here, Rpeak indicates that the order of speckle patterns, 4.3. Results
from good quality to poor quality, is h > a > g > f > d > e > c > b.
Fig. 6(a)–(c) show the SE, MIG, and MFFI of the eight speckle pat­ To present the displacements accurately, a series of translated images
terns, respectively. For the SDSPS of the MFFI in this work, the image of the eight speckle patterns was generated by applying the Fourier
was first transformed into a binary image by using Otsu’s method [31]. shifting method. The sub-pixel displacements applied in the x-direction
Subsequently, the MATLAB function “imopen” was used to locate the ranged from 0 to 1 pixels, corresponded to a shift of 0.05 pixels between

Fig. 4. Artificial speckle patterns and the corresponding histograms, (b) made by marker pen, (g) made by brush, and others are made by airbrush with different
colored backgrounds and speckles.

5
X. Hu et al. Measurement 173 (2021) 108618

Fig. 5. (a) Autocorrelation, (b) autocorrelation peak sharpness radius (Rpeak). Pattern index a–h corresponds to the eight speckle patterns in Fig. 4.

Fig. 6. (a) Shannon entropy (SE), (b) mean intensity gradient (MIG), and (c) multi-factor fusion index (MFFI) of the speckle patterns (a)–(h) in Fig. 4.

than 10− 6 .
According to Refs. [5,33], the mean bias error and standard devia­
tion error of measured displacements are defined as follows:
eu = umean − upre (20)
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√
√ 1 ∑ N
σu = √ (ui − umean )2 (21)
N − 1 i=1

where N is the number of grid points, upre is the pre-defined sub-pixel

∑
displacement, ui is the calculated displacement, umean = (1/N) Ni=1 ui is
the mean value of the N estimated displacements, eu is the mean bias
error of the calculated displacements, and σ u is the standard deviation
error of the calculated displacements.

4.3.1. Influence of noise and subset size

According to Section 3.2, the proposed parameter is based on the
rigid-body translation. In order to verify whether the influences of noise
Fig. 7. Ef of the speckle patterns (a)–(h) in Fig. 4. and subset size on parameter Ef could be ignored for evaluating the
entire speckle pattern (similar to MIG), the rigid-body translation of 0.25
two successive images [32]. The region denoted by a red rectangle as pixels was utilized to analysis the effect of noise and subset size on the
shown in Fig. 4(a) was analyzed. In total there were 100 × 100 grid mean bias error. Additive random Gaussian noises with mean value of 0,
points with a 5-pixel grid step in the reference image. The IC-GN algo­ variance of 0 to 3 × 10− 4 are added to the images which are normalized
rithm with a first-order shape function and a bicubic spline interpolation to 0–1, where the step size of noise variance is 3 × 10− 4 . The subset size
method was used to optimize the ZNSSD correlation criterion [5]. In this ranged from 15 × 15 pixels to 41 × 41 pixels with a step of 4 pixels.
work, the convergence condition was set to ensure that variation in the Fig. 8(a) shows the relationship between the mean bias error and various
norm of the incremental deformation parameter was equal to or less noises with a subset size of 41 × 41 pixels. Fig. 8(b) shows the

6
X. Hu et al. Measurement 173 (2021) 108618

Fig. 8. Mean bias error interrelated to (a) variance of noise (10− 4) and (b) subset size.

relationship between the mean bias error and subset size when the noise signifying that the mean bias error is robust to the subset size [34].
variance was 1 × 10− 4 . Therefore, the effect of noise and subset size on the speckle pattern
It can be concluded from Fig. 8 that (1) the mean bias error of the quality assessment can be ignored when the IC-GN algorithm is used,
displacements calculated by the IC-GN algorithm remained steady with and the proposed parameter Ef can be used to evaluate the quality of the
little fluctuation as the noise level increased. (2) The mean bias error entire speckle pattern.
also remained nearly unchanged, while the subset size increased

Fig. 9. (a1)–(a2) Mean bias errors of the eight speckle patterns for rigid-body translation images over a 0–1pixel displacement range using 21 × 21 pixels subset (left)
and 41 × 41 pixels subset (right). (b1)–(b2) standard deviation errors using 21 × 21 pixels subset (left) and 41 × 41 pixels subset (right).

7
X. Hu et al. Measurement 173 (2021) 108618

4.3.2. Error estimation of the translated image speckle pattern with a larger intensity gradient may perform less well
Fig. 9 shows the mean bias errors and standard deviation errors than a speckle pattern with a smaller intensity gradient because the
under each imposed sub-pixel displacement for the eight speckle pat­ speckle pattern with a larger intensity gradient may also have a larger
terns. A Gaussian noise variance of 1 × 10− 4 was added to the reference second-order derivative, and lead to larger values of Eqs. (13) or (17).
and translated images. These errors were computed by using the 21 × 21 As observed in Fig. 10, the corresponding mean values of the
and 41 × 41 pixel subsets. measured displacements of the speckle patterns (a) to (h) increasingly
As observed in Fig. 9, the subset size has little influence on the mean diverge from 0.25. The mean bias error, from small to large, is a < b < c
bias error. However, larger subset sizes can reduce the influence of noise < d < e < f < g < h, indicating that the performance of the speckle
on the standard deviation error. The results of DIC show good agreement patterns, from good to poor, is a > b > c > d > e > f > g > h. Similar
with the theoretical estimations, and the mean bias errors indicate that results can be obtained for the standard deviation error. The orders of
the order of speckle patterns, from good quality to poor quality, is a > b speckle pattern quality obtained by Ef and MIG are consistent with the
> c > d > e > f > g > h; this order agrees with that estimated by Ef. The experiment results. As shown in Figs. 5–7, the orders of speckle pattern
standard deviation errors indicate that the order of speckle patterns, quality assessed by parameters Rpeak , SE and MFFI are different to that
from good quality to poor quality, is a > f > b > c > e > h > d > g, which obtained by parameters Ef and MIG. The derivation of the proposed
agrees with the order estimated by the MIG. The above results indicate parameter Ef indicates that the mean bias error of DIC is related to the
that a speckle pattern that has a smaller mean bias error may have a first-order and second-order gray intensity gradients, and the Ef com­
larger standard deviation error. For example, speckle pattern (h) has the bines the first-order and second-order gray intensity gradients to eval­
largest mean bias error, but its standard deviation error is not the uate the quality of different speckle pattern. So, it can be used to
largest. Only speckle pattern (a) has both the smallest mean bias error optimize and design the speckle pattern more effectively by considering
and the smallest standard deviation error. Therefore, a speckle pattern the first-order and second-order gray intensity gradients. However, the
with a larger intensity gradient may perform less well than a speckle parameters Rpeak , SE and MFFI are based on the morphology and gray-
pattern with a smaller intensity gradient. Fig. 9 indicates that the level intensity which cannot reflect the principle of the mean bias
parameter Ef efficiently evaluates the speckle quality from the error. Combining with these experiment results, the parameters Ef and
perspective of mean bias error, and the MIG efficiently evaluates the MIG are more efficient in assessing the speckle pattern quality than other
speckle quality from the standard deviation error perspective. The three parameters from the perspectives of mean bias error and standard
speckle pattern that has both a smaller mean bias error and a smaller deviation error, respectively.
standard deviation error performs better in DIC measurements.
5. Discussion
4.3.3. Subset size selection based on MIG
As analyzed in Section 4.3.1, the mean bias error of the pure rigid- 5.1. Effectiveness of Ef
body translation is minimally influenced by the subset size and noise
when using the IC-GN algorithm, while the standard deviation error is As shown in Ref. [6] and Fig. 9, the interpolation bias is nominally
substantially related to the subset size. As observed from Ref. [19], the anti-symmetric relative to the position of 0.5 pixel, and the interpolation
same standard deviation error can be achieved for different speckle bias being sinusoidal, with minima at the positions of integer and 0.5
patterns by adjusting the subset size according to Eq. (5). As shown in pixel. In Ref. [6], the gradient ∇gx (xij ) of the translated image varies
Fig. 9, since speckle pattern (a) has the smallest standard deviation with τx , in order to get an invariant that not varies with τx for estimating
error, it should be used as the comparison image. According to Eq. (5) parameter Ef in this study, the ∇gx (xij ) is replaced by ∇fx (xij ) for
and the MIG values, if the subset size is decided to be 13 × 13 pixels for simplification. Therefore, the mean bias error is almost a sinusoidal
speckle pattern (a), the standard deviation error of speckle pattern (a) curve because the gradients in the translated image are related to τx and
with a subset size of 13 pixels is 0.0053 pixels. To obtain the same estimated at non-integer positions, while Eqs. (13) and (17) are not si­
standard deviation error, the equivalent subset size should be used for nusoidal curves because the gradients are estimated at integer positions
the seven other speckle patterns, with the size of 19 pixels, 27 pixels, 31 in the reference image. The parameters K1 and K2 in Eqs. (13) and (17)
pixels, 33 pixels, 43 pixels, 47 pixels, and 55 pixels, respectively, for are related to τx , which are the same for all the speckle patterns, and can
patterns f, b, c, e, h, d, and g. Fig. 10 shows the deviation between the be regarded as constants. The K1 and K2 are omitted during the deri­
measured displacement and that for the eight speckles with a 0.25 pixel vation of parameter Ef, therefore, the order of the values of Efs is
displacement. By comparison, we can observe that approximately the insensitive to the rigid-body translation. As observed in Figs. 9 and 10,
same standard deviation error was obtained by using equivalent subset the umean deviates from one side of the pre-defined displacement for each
sizes. The mean bias errors of all eight speckle patterns are almost speckle pattern and the order of speckle pattern quality shows good
consistent even when different subset sizes were used. The mean bias agreement with that estimated by Ef. Therefore, the parameter Ef can be
error order is a < b < c < d < e < f < g < h, which agrees with that shown used to evaluate the speckle pattern quality.
in Fig. 9. For the standard deviation error, the order is a < f < b < c < e The speckle size and density also can affect the measurement accu­
< h < d < g, which also agrees with that shown in Fig. 9. racy of DIC. However, different speckle size and density ultimately
The effectiveness of the proposed assessment parameter is demon­ provide different first-order and second gradients which are deeply
strated above by adding rigid-body translations to eight different speckle related the mean bias error. Speckle image with appropriate speckle size
patterns. The results indicate that the proposed parameter efficiently and density can reduce the errors. The parameter Ef shows that the mean
evaluates the speckle pattern quality from the mean bias error bias error is related to the first-order and second-order gradients. For
perspective and the MIG efficiently evaluates it from the standard de­ speckle size, very large speckle size will provide smaller first-order and
viation error perspective. However, we should note that the noise term second-order gradients which will lead to larger mean bias error and
of the mean bias error is ignored in these derivations and experiments larger standard deviation error, according to Ef, the mean bias error is
because the IC-GN algorithm is robust to noise. The accuracy and mainly affected by the smaller first-order gradients in this case [22];
calculation efficiency of IC-GN performs better than FA-GN [29]. very small speckle size will provide larger first-order and larger second-
Therefore, IC-GN is more suitable to DIC measurement and the proposed order gradients which will lead to larger mean bias error and smaller
assessment parameter based on IC-GN is reasonable. If the FA-GN al­ standard deviation error, similarly, the mean bias error is mainly
gorithm is used in DIC measurement, according to Eqs. (4), (13), and affected by the larger second-order gradients in this case [22]. There­
(17), a high-quality speckle pattern should have a small second-order fore, the speckle pattern with an appropriate speckle size and density
derivative, which agrees with the conclusion of Yu et al [22]. A

8
X. Hu et al. Measurement 173 (2021) 108618

Fig. 10. Deviation of the measured displacement for the speckle patterns a–h using a fixed 13 × 13 subset size (top left corner) and equivalent subset.

9
X. Hu et al. Measurement 173 (2021) 108618

should be recommended [15]. The efficiency of parameter Ef on extreme Acknowledgments

cases of speckle size and density is worth for further study and then it
can be used to optimize the speckle size and density for patterns. This work was supported by the National Natural Science Foundation
of China (Grant No. 51605059), Nation Key Research and Development
Program (Grant Nos. 2016lYFE0113600, 2018YFB2001400), and
5.2. Total error assessment Fundamental Research Fund for the Central Universities (Grant No.
2019CDJGFJX003).
From the numerical experiments, the proposed parameter Ef is effi­
cient in assessing speckle pattern from the perspective of mean bias error References
and the parameter MIG is efficient in assessing speckle pattern from the
perspective of standard deviation error. Therefore, it seems that the Ef [1] K. Muniandy, Z. Mohamad Ariff, A. Abu Bakar, Digital image correlation utilization
in measuring displacement and strain during plastic film blowing process: a
and MIG should be combined to evaluate the total error of DIC mea­
feasibility study, Meas. J. Int. Meas. Confed. 136 (2019) 487–500, https://doi.org/
surement. The root-mean-square error (RMSE) is widely used to eval­ 10.1016/j.measurement.2018.12.093.
uate the total error of displacements as shown in Eq. (22) [15]. [2] N.K. Raghuwanshi, A. Parey, Experimental measurement of spur gear mesh
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ stiffness using digital image correlation technique, Meas. J. Int. Meas. Confed. 111
(2017) 93–104, https://doi.org/10.1016/j.measurement.2017.07.034.
RMSE = e2u + σ 2u (22) [3] Y.L. Dong, B. Pan, A review of speckle pattern fabrication and assessment for
digital image correlation, Exp. Mech. 57 (2017) 1161–1181, https://doi.org/
where eu is positively correlated with Ef, σ u is positively correlated with 10.1007/s11340-017-0283-1.
[4] G. Stoilov, V. Kavardzhikov, D. Pashkouleva, A comparative study of random
1/MIG. patterns for digital image correlation, J. Theor. Appl. Mech. 42 (2012) 55–66,
It is noted here that the RMSE is related to the displacement τx and https://doi.org/10.2478/v10254-012-0008-x.
the standard deviation of noise σ. As shown in Figs. 6 and 7, the orders of [5] X. Shao, X. Dai, X. He, Noise robustness and parallel computation of the inverse
compositional Gauss-Newton algorithm in digital image correlation, Opt. Lasers
speckle pattern quality of speckle patterns a, b, c, d, and g are the same Eng. 71 (2015) 9–19, https://doi.org/10.1016/j.optlaseng.2015.03.005.
from both the parameters Ef and MIG. Then the order of speckle pattern [6] Y.Q. Wang, M.A. Sutton, H.A. Bruck, H.W. Schreier, Quantitative error assessment
quality evaluated by RMSE will be consistent with Ef or MIG. On the in pattern matching: effects of intensity pattern noise, interpolation, strain and
image contrast on motion measurements, Strain 45 (2009) 160–178, https://doi.
contrary, for speckle patterns d, e, and f, the orders of speckle pattern
org/10.1111/j.1475-1305.2008.00592.x.
quality evaluated by Ef and MIG are different, and the standard devia­ [7] B. Pan, H. Xie, Z. Wang, K. Qian, Z. Wang, Study on subset size selection in digital
tion error is related to the unknown noise. Therefore, the influence de­ image correlation for speckle patterns, Opt. Express 16 (2008) 7037–7048, https://
doi.org/10.1364/oe.16.007037.
grees of parameters τx and σ on RMSE are unknown that the order of
[8] Q. Zhan, Y. Yuan, X. Fan, J. Huang, C. Xiong, F. Yuan, Digital image correlation
speckle pattern quality evaluated by RMSE might be different under involves an inverse problem: a regularization scheme based on subset size
different displacements and noise levels. constraint, Opt. Lasers Eng. 81 (2016) 54–62, https://doi.org/10.1016/j.
optlaseng.2016.01.007.
[9] Z. Chen, C. Quan, F. Zhu, X. He, A method to transfer speckle patterns for digital
6. Conclusions image correlation, Meas. Sci. Technol. (2015), https://doi.org/10.1088/0957-
0233/26/9/095201.
[10] J. Zhang, A. Sweedy, F. Gitzhofer, G. Baroud, A novel method for repeatedly
In DIC, speckle patterning is vital for displacement measurement. To generating speckle patterns used in digital image correlation, Opt. Lasers Eng. 100
accurately assess the performances of different speckle patterns, a global (2018) 259–266, https://doi.org/10.1016/j.optlaseng.2017.09.012.
assessment parameter Ef that combines the first-order gradient and [11] F. Sur, B. Blaysat, M. Grédiac, Rendering deformed speckle images with a Boolean
model, J. Math. Imaging Vis. 60 (2018) 634–650, https://doi.org/10.1007/
second-order gradient is proposed. A theoretical deviation of the
s10851-017-0779-4.
parameter is given, and eight specimens with different speckle patterns [12] G. Gu, A comparative study of random speckle pattern simulation models in digital
were used to verify the efficiency. The proposed parameter was image correlation, Optik (Stuttg) 126 (2015) 3713–3716, https://doi.org/
10.1016/j.ijleo.2015.08.271.
compared to other existing parameters, and the speckle pattern quality
[13] Y. Dong, H. Kakisawa, Y. Kagawa, Development of microscale pattern for digital
orders that were evaluated by different parameters were obtained. Nu­ image correlation up to 1400 1 C, Opt. Lasers Eng. 68 (2015) 7–15, https://doi.
merical experiments with rigid-body translations were performed on the org/10.1016/j.optlaseng.2014.12.003.
eight speckle patterns. The results show that the mean bias error is [14] G. Lionello, C. Sirieix, M. Baleani, An effective procedure to create a speckle
pattern on biological soft tissue for digital image correlation measurements,
related to the second derivative of gray intensity and it is substantially J. Mech. Behav. Biomed. Mater. 39 (2014) 1–8.
affected by the parameter Ef. A speckle pattern with smaller Ef produced [15] Z. Chen, X. Shao, X. Xu, X. He, Optimized digital speckle patterns for digital image
a smaller mean bias error. For speckle optimization and preparation, correlation by consideration of both accuracy and efficiency, Appl. Opt. 57 (2018)
884–893, https://doi.org/10.1364/ao.57.000884.
both the first-order and second-order gradients should be considered. [16] Y. Su, Q. Zhang, Z. Gao, Statistical model for speckle pattern optimization, Opt.
These results also indicate that the MIG is efficient in evaluating the Express. 25 (2017) 30259, https://doi.org/10.1364/oe.25.030259.
speckle pattern quality from the perspective of standard deviation. Thus, [17] Y. Su, Z. Gao, Z. Fang, Y. Liu, Y. Wang, Q. Zhang, S. Wu, Theoretical analysis on
performance of digital speckle pattern: uniqueness, accuracy, precision, and spatial
operators should combine the Ef parameter and the MIG parameter to resolution, Opt. Express. 27 (2019) 22439–22474, https://doi.org/10.1364/
evaluate the accuracy and precision of DIC measurements using oe.27.022439.
different speckle patterns in DIC experiments. [18] G.F. Bomarito, J.D. Hochhalter, T.J. Ruggles, A.H. Cannon, Increasing accuracy
and precision of digital image correlation through pattern optimization, Opt.
Lasers Eng. 91 (2017) 73–85, https://doi.org/10.1016/j.optlaseng.2016.11.005.
CRediT authorship contribution statement [19] B. Pan, Z. Lu, H. Xie, Mean intensity gradient: An effective global parameter for
quality assessment of the speckle patterns used in digital image correlation, Opt.
Lasers Eng. 48 (2010) 469–477, https://doi.org/10.1016/j.optlaseng.2009.08.010.
Xiaoliang Hu: Investigation, Data curation, Software, Writing - [20] T. Hua, H. Xie, S. Wang, Z. Hu, P. Chen, Q. Zhang, Evaluation of the quality of a
original draft. Zhijiang Xie: Visualization, Investigation, Validation. Fei speckle pattern in the digital image correlation method by mean subset fluctuation,
Liu: Conceptualization, Methodology, Supervision, Writing - review & Opt. Laser Technol. 43 (2011) 9–13, https://doi.org/10.1016/j.
optlastec.2010.04.010.
editing. [21] X. Zhao, Y. Wen, J. Zhao, D. Zhao, Study of the quality of wood texture patterns in
digital image correlation, Optik (Stuttg) 171 (2018) 370–376, https://doi.org/
10.1016/j.ijleo.2018.06.017.
Declaration of Competing Interest [22] H. Yu, R. Guo, H. Xia, F. Yan, Y. Zhang, T. He, Application of the mean intensity of
the second derivative in evaluating the speckle patterns in digital image
correlation, Opt. Lasers Eng. 60 (2014) 32–37, https://doi.org/10.1016/j.
The authors declare that they have no known competing financial optlaseng.2014.03.015.
interests or personal relationships that could have appeared to influence [23] X.Y. Liu, R.L. Li, H.W. Zhao, T.H. Cheng, G.J. Cui, Q.C. Tan, G.W. Meng, Quality
assessment of speckle patterns for digital image correlation by Shannon entropy,
the work reported in this paper.

10
X. Hu et al. Measurement 173 (2021) 108618

Optik (Stuttg) 126 (2015) 4206–4211, https://doi.org/10.1016/j. [29] Y. Gao, T. Cheng, Y. Su, X. Xu, Y. Zhang, Q. Zhang, High-efficiency and high-
ijleo.2015.08.034. accuracy digital image correlation for three-dimensional measurement, Opt. Lasers
[24] Y. Su, Q. Zhang, X. Xu, Z. Gao, Quality assessment of speckle patterns for DIC by Eng. 65 (2014) 73–80, https://doi.org/10.1016/j.optlaseng.2014.05.013.
consideration of both systematic errors and random errors, Opt. Lasers Eng. 86 [30] S. Bossuyt, Optimized patterns for digital image correlation, Conf. Proc. Soc. Exp.
(2016) 132–142, https://doi.org/10.1016/j.optlaseng.2016.05.019. Mech. Ser. 3 (2013) 239–248, https://doi.org/10.1007/978-1-4614-4235-6_34.
[25] J. Park, S. Yoon, T.H. Kwon, K. Park, Assessment of speckle-pattern quality in [31] N. Otsu, Threshold selection method from gray-level histograms, IEEE Trans. Syst.
digital image correlation based on gray intensity and speckle morphology, Opt. Man Cybern. SMC-9 (1979) 62–66, https://doi.org/10.1109/tsmc.1979.4310076.
Lasers Eng. 91 (2017) 62–72, https://doi.org/10.1016/j.optlaseng.2016.11.001. [32] B. Pan, K. Li, W. Tong, Fast, robust and accurate digital image correlation
[26] J. Song, J. Yang, F. Liu, K. Lu, Quality assessment of laser speckle patterns for calculation without redundant computations, Exp. Mech. 53 (2013) 1277–1289,
digital image correlation by a Multi-Factor Fusion Index, Opt. Lasers Eng. 124 https://doi.org/10.1007/s11340-013-9717-6.
(2020), https://doi.org/10.1016/j.optlaseng.2019.105822. [33] S. Simončič, P. Podržaj, An improved digital image correlation calculation in the
[27] D. Lecompte, A. Smits, S. Bossuyt, H. Sol, J. Vantomme, D. Van Hemelrijck, A. case of substantial lighting variation, Exp. Mech. 57 (2017) 743–753, https://doi.
M. Habraken, Quality assessment of speckle patterns for DIC, Opt. Lasers Eng. 44 org/10.1007/s11340-017-0265-3.
(2006) 1132–1145, https://doi.org/10.1016/j.optlaseng.2005.10.004. [34] S. Yaofeng, J.H.L. Pang, Study of optimal subset size in digital image correlation of
[28] G. Crammond, S.W. Boyd, J.M. Dulieu-Barton, Speckle pattern quality assessment speckle pattern images, Opt. Lasers Eng. 45 (2007) 967–974, https://doi.org/
for digital image correlation, Opt. Lasers Eng. 51 (2013) 1368–1378, https://doi. 10.1016/j.optlaseng.2007.01.012.
org/10.1016/j.optlaseng.2013.03.014.

11