You are on page 1of 17

Vol. 25, No.

24 | 27 Nov 2017 | OPTICS EXPRESS 30259

Statistical model for speckle pattern


optimization
YONG S U , Q INGCHUAN Z HANG , * AND Z EREN G AO
CAS Key Laboratory of Mechanical Behavior and Design of Materials, Department of Modern Mechanics,
University of Science and Technology of China, Hefei 230027, China
* zhangqc@ustc.edu.cn

Abstract: Image registration is the key technique of optical metrologies such as digital image
correlation (DIC), particle image velocimetry (PIV), and speckle metrology. Its performance
depends critically on the quality of image pattern, and thus pattern optimization attracts extensive
attention. In this article, a statistical model is built to optimize speckle patterns that are composed
of randomly positioned speckles. It is found that the process of speckle pattern generation
is essentially a filtered Poisson process. The dependence of measurement errors (including
systematic errors, random errors, and overall errors) upon speckle pattern generation parameters
is characterized analytically. By minimizing the errors, formulas of the optimal speckle radius are
presented. Although the primary motivation is from the field of DIC, we believed that scholars in
other optical measurement communities, such as PIV and speckle metrology, will benefit from
these discussions.
© 2017 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
OCIS codes: (100.2000) Digital image processing; (120.0120) Instrumentation, measurement, and metrology; (120.3940)
Metrology.

References and links


1. M. A. Sutton, J. J. Orteu, and H. Schreier, Image Correlation for Shape, Motion and Deformation Measurements:
Basic Concepts, Theory and Applications (Springer Science & Business Media, 2009).
2. M. Raffel, C. E. Willert, S. Wereley, and J. Kompenhans, Particle Image Velocimetry: a Practical Guide (Springer,
2013).
3. M. Sjödahl and L. R. Benckert, “Systematic and random errors in electronic speckle photography,” Appl. Opt. 33(31),
7461–7471 (1994).
4. M. Chakrabarti, M. L. Jakobsen, and S. G. Hanson, “Speckle-based spectrometer,” Opt. Lett. 40(14), 3264–3267
(2015).
5. R. Szeliski, Computer Vision: Algorithms and Applications (Springer Science & Business Media, 2010).
6. B. Zitová and J. Flusser, “Image registration methods: a survey,” Image Vis. Comput. 21(11), 977–1000 (2003).
7. D. Lecompte, A. Smits, S. Bossuyt, H. Sol, J. Vantomme, D. Van Hemelrijck, and A. M. Habraken, “Quality
assessment of speckle patterns for digital image correlation,” Opt. Lasers Eng. 44(11), 1132–1145 (2006).
8. Y. Su, Q. Zhang, X. Xu, and Z. Gao, “Quality assessment of speckle patterns for DIC by consideration of both
systematic errors and random errors,” Opt. Lasers Eng. 86, 132–142 (2016).
9. L. Luu, Z. Wang, M. Vo, T. Hoang, and J. Ma, “Accuracy enhancement of digital image correlation with B-spline
interpolation,” Opt. Lett. 36(16), 3070–3072 (2011).
10. Y. Su, Q. Zhang, Z. Gao, and X. Xu, “Noise-induced bias for convolution-based interpolation in digital image
correlation,” Opt. Express 24(2), 1175–1195 (2016).
11. H. W. Schreier, J. R. Braasch, and M. A. Sutton, “Systematic errors in digital image correlation caused by intensity
interpolation,” Opt. Eng. 39(11), 2915–2921 (2000).
12. T. Roesgen, “Optimal subpixel interpolation in particle image velocimetry,” Exp. Fluids 35(3), 252–256 (2003).
13. Y. Su, Q. Zhang, Z. Gao, X. Xu, and X. Wu, “Fourier-based interpolation bias prediction in digital image correlation,”
Opt. Express 23(15), 19242–19260 (2015).
14. Y. Su, Q. Zhang, X. Xu, Z. Gao, and S. Wu, “Interpolation bias for the inverse compositional Gauss-Newton algorithm
in digital image correlation,” Opt. Lasers Eng. 100, 267–278 (2018).
15. Y. Wang, M. Sutton, H. Bruck, and H. Schreier, “Quantitative error assessment in pattern matching: effects of intensity
pattern noise, interpolation, strain and image contrast on motion measurements,” Strain 45(2), 160–178 (2009).
16. B. Blaysat, M. Grédiac, and F. Sur, “On the propagation of camera sensor noise to displacement maps obtained by
DIC - an experimental study,” Exp. Mech. 56(6), 919–944 (2016).
17. B. Pan, H. Xie, Z. Wang, K. Qian, and Z. Wang, “Study on subset size selection in digital image correlation for
speckle patterns,” Opt. Express 16(10), 7037–7048 (2008).

#307453 https://doi.org/10.1364/OE.25.030259
Journal © 2017 Received 19 Sep 2017; revised 1 Nov 2017; accepted 1 Nov 2017; published 17 Nov 2017
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30260

18. X. Shao, X. Dai, and X. He, “Noise robustness and parallel computation of the inverse compositional Gauss-Newton
algorithm in digital image correlation,” Opt. Lasers Eng. 71, 9–19 (2015).
19. W. Chen, Z. Jiang, L. Tang, Y. Liu, and Z. Liu, “Equal noise resistance of two mainstream iterative sub-pixel
registration algorithms in digital image correlation,” Exp. Mech. 57(6), 979–996 (2017).
20. P. Zhou and K. E. Goodson, “Subpixel displacement and deformation gradient measurement using digital image/speckle
correlation (DISC),” Opt. Eng. 40(8), 1613–1620 (2001).
21. D. Lecompte, H. Sol, J. Vantomme, and A. Habraken, “Analysis of speckle patterns for deformation measurements
by digital image correlation,” Proc. SPIE 6341, 63410 (2006).
22. T. Hua, H. Xie, S. Wang, Z. Hu, P. Chen, and Q. Zhang, “Evaluation of the quality of a speckle pattern in the digital
image correlation method by mean subset fluctuation,” Opt. Laser Technol. 43(1), 9–13 (2011).
23. Z. Hu, T. Xu, H. Luo, R. Z. Gan, and H. Lu, “Measurement of thickness and profile of a transparent material using
fluorescent stereo microscopy,” Opt. Express 24(26), 29822–29829 (2016).
24. P. Mazzoleni, F. Matta, E. Zappa, M. A. Sutton, and A. Cigada, “Gaussian pre-filtering for uncertainty minimization
in digital image correlation using numerically-designed speckle patterns,” Opt. Lasers Eng. 66, 19–33 (2015).
25. G. Bomarito, J. Hochhalter, T. Ruggles, and A. Cannon, “Increasing accuracy and precision of digital image
correlation through pattern optimization,” Opt. Lasers Eng. 91, 73–85 (2017).
26. Z. Chen, C. Quan, F. Zhu, and X. He, “A method to transfer speckle patterns for digital image correlation,” Meas. Sci.
Technol. 26(9), 095201 (2015).
27. A. Lavatelli and E. Zappa, “A displacement uncertainty model for 2-D DIC measurement under motion blur
conditions,” IEEE Trans. Instrum. Meas. 66(3), 451–459 (2017).
28. J. Westerweel, “Fundamentals of digital particle image velocimetry,” Meas. Sci. Technol. 8(12), 1379 (1997).
29. A. Papoulis and S. U. Pillai, Probability, Random Variables, and Stochastic Processes (Tata McGraw-Hill Education,
2002).
30. Y. Wang, P. Lava, P. Reu, and D. Debruyne, “Theoretical analysis on the measurement errors of local 2D DIC: part I
temporal and spatial uncertainty quantification of displacement measurements,” Strain 52(2), 110–128 (2015).
31. S. O. Rice, “Mathematical analysis of random noise,” Bell Syst. Tech. J. 23(3), 282–332 (1944).
32. B. Pan, K. Qian, H. Xie, and A. Asundi, “Two-dimensional digital image correlation for in-plane displacement and
strain measurement: a review,” Meas. Sci. Technol. 20(6), 062001 (2009).
33. P. Thevenaz, T. Blu, and M. Unser, “Interpolation revisited,” IEEE Trans. Med. Imaging 19(7), 739–758 (2000).
34. H. A. Bruck, S. R. McNeill, M. A. Sutton, and W. H. Peters, “Digital image correlation using Newton-Raphson
method of partial differential correction,” Exp. Mech. 29(3), 261–267 (1989).
35. S. Baker and I. Matthews, “Lucas-Kanade 20 years on: a unifying framework,” Int. J. Comput. Vision 56(3), 221–255
(2004).
36. B. Pan, K. Li, and W. Tong, “Fast, robust and accurate digital image correlation calculation without redundant
computations,” Exp. Mech. 53(7), 1277–1289 (2013).
37. Y. Gao, T. Cheng, Y. Su, X. Xu, Y. Zhang, and Q. Zhang, “High-efficiency and high-accuracy digital image correlation
for three-dimensional measurement,” Opt. Lasers Eng. 65, 73–80 (2015).
38. B. Kumar and S. C. D. Roy, “Design of digital differentiators for low frequencies,” Proc. IEEE 76(3), 287–289 (1988).
39. Y. Su, “Optimal speckle radius for image registration,” figshare (2017) [retrieved 01 Nov 2017],
http://dx.doi.org/10.6084/m9.figshare.5558542.v1.
40. W. Tong, “Subpixel image registration with reduced bias,” Opt. Lett. 36(5), 763–765 (2011).
41. Z. Gao, X. Xu, Y. Su, and Q. Zhang, “Experimental analysis of image noise and interpolation bias in digital image
correlation,” Opt. Lasers Eng. 81, 46–53 (2016).
42. W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Numerical Recipes: The Art of Scientific Computing
(Cambridge University, 2007).

1. Introduction
Image registration is the key technique of optical metrologies such as digital image correlation
(DIC) [1], particle image velocimetry (PIV) [2], speckle metrology [3,4], and motion estimation [5].
It registers the same physical points in two or more images of the same scene captured at different
times, from different sensors, or from different viewpoints [6]. The metrological performance of
image registration depends critically on the quality of image pattern [7]. Two issues then arise:
(1) how to assess a pattern; (2) how to generate a good pattern.
The first issue, pattern assessment, has been thoroughly studied [8]. Since a good pattern should
correspond to a small measurement error, the overall error is a valid pattern quality measure [9].
The overall error is composed of systematic error and random error [10]. The systematic error is
caused by imperfect interpolation [11,12]; it is largely determined by the interpolation bias kernel
and the image power spectrum [13, 14]. The random error stems from sensor noise [15, 16]; it
can be accurately characterized by the ratio of the variance of sensor noise and the sum of square
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30261

of subset intensity gradients [17–19].


The second issue, pattern optimization, has attracted extensive attention [20–23]. However,
finding the optimal pattern is difficult, and thus patterns of specific type are generally optimized
instead [24,25]. In practical situations, speckle patterns which are composed of random positioned
speckles are widely used because of their simplicity, practicability, and effectiveness [26, 27].
The speckles have physical meanings sometimes, for example fluorescent particles of fluorescent
stereo microscopy and tracer particles of PIV [23, 28]. These patterns can be characterized by
several generation parameters such as speckle size and speckle coverage, and hence pattern
optimization can be implemented by optimizing these generation parameters [21]. A number of
literatures have been devoted to optimizing the generation parameters: Zhou et al. proposed that
the optimal size of Gaussian speckle should lie within 2∼5 pixels [20]; Lecompte et al. showed
that when a subset 15 is chosen the optimal choice for speckle diameter is 5 pixels and the speckle
coverage should lie between 40%∼70% [21]; Sutton et al. suggested that to ensure a reasonable
amount of over-sampling the optimal size should be 3∼6 pixels [1]; Hua et al. proposed that the
speckle size should be from 2 to 4 pixels [22].
Nowadays, the main problem in the subject of speckle pattern optimization is the lack of a
theoretical model. All existing achievements are attained using the numerical methods [20–22].
The numerical methods can only determine the optimal speckle pattern generation parameters in
some particular cases, for it is impossible to conduct a numerical experiment for every possible
parameter. Furthermore, the numerical methods are incapable of providing a deep physical insight.
Thus, there is a strong need to build a theoretical model, and it is the objective of this article.
In this article, a statistical model for speckle pattern optimization is built and validated.
The theoretical analyses are divided into three gradually progressive steps: firstly, the statistics
(including expectation, variance, and power spectrum) of speckle patterns are derived in Section
2; then, the measurement errors (including systematic errors, random errors, and overall errors)
are derived in Section 3; finally, the pattern quality measure is derived and formulas of optimal
speckle radius are presented in Section 4. The proposed theoretical model is confirmed by
numerical experiments in Section 5. The advantages and limits of this research are discussed
in Section 6 and the conclusions are drawn in Section 7. The primary motivation of this article
is to establish a speckle pattern generation standard for DIC. However, since the mathematical
derivations are general, it is believed that these discussions are beneficial to scholars in other
optical measurement communities, such as PIV and speckle metrology.

2. Statistic of speckle patterns


Image patterns play a crucial role in the metrological performance [7]. In practical situations,
speckle patterns which are composed of random positioned speckles are widely used [20–23, 26–
28]. These patterns can be mathematically modeled as
N
Õ
f(x, y) = ψ(x − xi, y − yi ), (1)
i=1

where N is the total number of speckles, ψ(x, y) is the speckle profile, and (xi, yi ) is the position of
the ith speckle. Following the convention in [29], random variables, such as xi and yi , are written
in boldface letters. As the speckle patterns should be isotropic [1], the speckle profile ψ(x, y) is
assumed to be isotropic. The N speckles are randomly scattered over a region [−L, L] × [−L, L],
and the positions of every two speckles are independent. Since the speckle positions (xi, yi ) are
random variables, identical speckle pattern generation parameters will generate different speckle
patterns, as illustrated in Fig. 1. This randomness is designated as spatial randomness, in contrast
with temporal randomness, which is caused by the randomness of sensor noise [30].
To register the correspondence, a subset is generally chosen around the point of interest. Since
the subset is generally much smaller than the speckle region, it is reasonable to assume that, from
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30262

(a) y (b) y

x x

2L

Fig. 1. A schematic of speckle patterns which are composed of random positioned speckles.
These speckles are randomly scattered over a region, and thus the speckle positions are
random variables. Due to this spatial randomness, patterns shown in (a) and (b) are different,
though the generation parameters are identical.

a performance perspective, patterns described by Eq. (1) are equivalent to patterns filling the
whole plane R2 with a speckle density λ = N/4L 2 . Namely, we assume L → ∞. These speckle
patterns can be rewritten as

Õ
f(x, y) = ψ(x − xi, y − yi ). (2)
i=−∞

The question then arises as to character the speckle pattern. One contribution of this work is the
find of the relationship between speckle pattern and filtered Poisson process.
If points are randomly scattered over the plane R2 with a constant density λ, and each point is
stochastically independent of all the other points, then this is a Poisson process [29], which can
be described using the Poisson impulses

Õ
z(x, y) = δ(x − xi, y − yi ), (3)
i=−∞

where δ(x, y) is the Dirac delta function, and (xi, yi ) is the position of the ith points . This process
is named after Poisson because it relates to the Poisson distribution [29].
Convolving the Poisson impulses with a filter ψ(x, y) leads to a filtered Poisson process [29],
which can be given by

Õ
f(x, y) = ψ ⊗ z(x, y) = ψ(x − xi, y − yi ). (4)
i=−∞

A comparison between Eq. (2) and Eq. (4) implies that the process of speckle pattern generation is
essentially a filtered Poisson process. Fortunately, the filtered Poisson process has been thoroughly
studied already [31]; the achievements can be exploited directly. The expectation and the variance
of a filtered Poisson process can be given by the Campbell’s theorem [31]
∫ ∞∫ ∞ ∫ ∞∫ ∞
E{f(x, y)} = λ ψ(ξ, ζ) dξdζ, Var(f(x, y)) = λ ψ 2 (ξ, ζ) dξdζ . (5)
−∞ −∞ −∞ −∞

The power spectrum of a filtered Poisson process is [29]


S f f (ωx, ωy ) = 4π 2 λ2 ψ̂ 2 (0, 0)δ(ωx, ωy ) + λ| ψ̂(ωx, ωy )| 2, (6)
where ψ̂(ωx, ωy ) is the Fourier transform of ψ(x, y).
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30263

3. Measurement errors
Measurement errors of pure translation are frequently applied to assessing the pattern quality [20].
For a pattern f (x, y), a target pattern f (x − u0, y) is produced by shifting the original pattern by u0
units along the x axis. Then, the original pattern and the target pattern are sampled, producing a
reference image f [m, n] and a target image g[m, n] respectively. In practice, images are inevitably
contaminated by sensor noise. If the noise in the reference image is w f [m, n] and the noise in the
target image is wg [m, n], the captured images are

f[m, n] = f (m, n) + w f [m, n], g[m, n] = f (m − u0, n) + wg [m, n]. (7)

A area-based registration strategy is commonly employed to estimate the imposed displacement


u0 [6]. A subset with size M is chosen and then specific similarity measure is optimized [32].
Without a loss of generality, the subset center is assumed to be at the origin. If the sum of squared
difference (SSD) criterion is utilized, the estimated displacement is given by
S
Õ S
Õ
u = arg min {f[m, n] − g(m + u, n)}2 , (8)
m=−S n=−S

where S = (M − 1)/2 denotes the half of subset size, and g(m + u, n) denotes the subpixel intensity
of target image, which can be estimated by interpolation [33]. Equation (8) is generally solved by
numerical optimization algorithms such as Newton-Raphson method [34], Levenberg-Marquardt
method, and Gauss-Newton method. In this article, the inverse compositional Gauss-Newton
(IC-GN) algorithm [35], which is the most popular subpixel registration algorithm in the filed
of DIC for its high-accuracy and high-efficiency [36, 37], is utilized to estimate the actual
displacement u0 .
The estimated displacement u is not equal to the actual displacement u0 but contains a
measurement error ∆u = u − u0 . ∆u is a random variable due to temporal randomness and can
be divided into two parts: a systematic error eb and a random error en , where
p
eb = E {∆u | f (x, y)} , en = Var {∆u | f (x, y)}. (9)

The subscript ‘b’ denotes ‘bias’ and ‘n’ denotes ‘noise’. It is worth noting that the random error
en is caused by sensor noise, not by speckle generation. If the spatial randomness is considered,
both systematic errors and random errors become random variables. The major objective of this
section is to derive the spatial expectations of eb and en .

3.1. Systematic errors


In this section, the spatial expectations of systematic errors are derived. If the inverse compositional
Gauss-Newton algorithm is utilized, the systematic error eb of a given pattern f (x, y) can be
approximated as [14]

eb ≈ Cb sin 2πu0,
∫π ∫π
Eb (ωx, ωy )| fˆ(ωx, ωy )| 2 dωx dωy
Cb = ∫ π ∫ π −π −π , (10)
−π −π x
ω d(ω
ˆ x, ωy )ϕ̂(ωx, ωy )| fˆ(ωx, ωy )| 2 dωx dωy
Eb (ωx, ωy ) = d(ω
ˆ x, ωy ) ϕ̂(ωx − 2π, ωy ) − ϕ̂(ωx + 2π, ωy ) .
 

where Cb is the amplitude of the systematic error, u0 is the actual displacement, Eb (ωx, ωy ) is
the interpolation bias kernel, fˆ(ωx, ωy ) is the spectrum of the speckle pattern, ϕ̂(ωx, ωy ) is the
interpolation transfer function of interpolation algorithm for target image, and d(ωˆ x, ωy ) is the
magnitude frequency response of gradient estimator for reference image.
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30264

In the context of speckle pattern generation, the pattern spectrum f̂(ωx, ωy ) becomes a stochastic
process. Accordingly, the systematic error eb becomes a random variable. Our aim is to estimate
the expectation of eb . Since the process of speckle pattern generation is essentially a filtered
Poisson process, it is rational to substitute | f̂(ωx, ωy )| 2 with the power spectrum of a filtered
Poisson process, thus
∫π ∫π
Eb (ωx, ωy )S f f (ωx, ωy ) dωx dωy
E{Cb } ≈ ∫ π ∫ π −π −π , (11)
−π −π x
ω d(ωx, ωy )ϕ̂(ωx, ωy )S f f (ωx, ωy ) dωx dωy
ˆ

where S f f (ωx, ωy ), given by Eq. (6), is the power spectrum of a filtered Poisson process. The
interpolation bias kernel Eb (ωx, ωy ) is a high-pass filter and equals zero at ωx = ωy = 0 [13]. A
further simplification yields

E{eb } ≈ E{Cb } sin 2πu0,


∫π ∫π
−π −π b
E (ωx, ωy )| ψ̂(ωx, ωy )| 2 dωx dωy
E{Cb } ≈ ∫ π ∫ π , (12)
−π −π
ωx d(ω
ˆ x, ωy )ϕ̂(ωx, ωy )| ψ̂(ωx, ωy )| 2 dωx dωy
Eb (ωx, ωy ) = d(ω
ˆ x, ωy ) ϕ̂(ωx − 2π, ωy ) − ϕ̂(ωx + 2π, ωy ) .
 

where ψ̂(ωx, ωy ) is the spectrum of an individual speckle. Equation (12) represents an estimation
of the spatial expectation of the systematic error.

3.2. Random errors


In this section, the spatial expectations of random errors are derived. If the inverse compositional
Gauss-Newton algorithm is utilized, the random error en of a given speckle pattern f (x, y) can
be given by [18] v
σf2 + σg2
t
en = Í S ÍS 2
, (13)
m=−S n=−S fx [m, n]
where σf is the standard deviation of sensor noise of reference image, σg is the standard deviation
of sensor noise of target image, S is the half of subset size, and fx [m, n] is the image gradient
along the x axis.
In the context of speckle pattern generation, the speckle pattern f(x, y) becomes a stochastic
process. Accordingly, the random error en becomes a random variable. Our aim is to estimate the
spatial expectation of en . The random error given by Eq. (13) can be rewritten as
σ
en = √ , (14)
Q

where σ = (σf2 + σg2 )1/2 , and Q = Sm=−S Sn=−S fx2 [m, n] is the sum of square of subset intensity
Í Í

gradients. Deriving E{en } straightforwardly seems difficult, and thus an estimate is presented
instead. This estimation is based on a theorem in [29]: the expectation of a random variable
y = Φ(x) can be estimated by approximating Φ(x) by a parabola, yielding
1
E{Φ(x)} ≈ Φ (E{x}) + Var(x)Φ00(E{x}). (15)
2
The second term, Var(x)Φ00(E{x})/2, acts as a correction [29]. In our case, Φ(x) = σx −1/2 , and
thus the expectation of the random error can be approximated by

σ
 
3 Var(Q)
E{en } ≈ p 1+ . (16)
E{Q} 8 E{Q}2
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30265

Hence, the issue turns to investigate the expectation and the variance of Q.
The expectation of Q can be derived as follows. The pattern gradient is generally evaluated by
convolving the pattern with a gradient operator d [5, 14], namely,

fx (x, y) = (d ⊗ f)(x, y) = (d ⊗ ψ ⊗ z)(x, y). (17)

Obviously, the pattern gradient fx (x, y) is a filtered Poisson process with filter d ⊗ ψ. Consequently,
its expectation and variance can be given by the Campbell’s theorem [see Eq. (5)]. As the speckle
patterns should be isotropic and homogeneous in a statistical sense [1], the expectation of the
pattern gradient should be zero. A combination of Parseval’s theorem and Campbell’s theorem
yields that the expectation of squared pattern gradient is
λ
∫ ∞∫ ∞
q = E{fx2 (x, y)} = 2 ˆ x, ωy )| 2 | ψ̂(ωx, ωy )| 2 dωx dωy .
| d(ω (18)
4π −∞ −∞
Since filtered Poisson process is strict-sense stationary [29], all points have the same statistics,
and thus the expectation of Q can be expressed as
S
Õ S
Õ
E{Q} = E fx2 [m, n] = M 2 q,

(19)
m=−S n=−S

where M = 2S + 1 is the subset size.


The variance of Q is difficult to derive. However, as the variance acts as a correction, a good
estimate seems enough. The sum of the squares of k independent standard normal random
variables satisfies a chi-squared distribution χk2 [29]. In practice, there are generally hundreds of
points in a subset. Recalling that the gradient of each point is a random variable with zero mean
and variance q, it is assumed that q−1/2 Q approximately satisfies a chi-squared distribution χM 2 ,
2
then the variance [29]
Var (Q) ≈ 2M 2 q. (20)
A combination of Eq. (16), Eq. (19), and Eq. (20) yields
σ
 
3
E{en } ≈ √ 1 + . (21)
M q 4M 2 q
Equation (21) represents an estimate of the spatial expectation of the random error.

3.3. Overall errors


In this section, the spatial expectations of overall errors are derived. The overall errors are
generally characterized by the mean squared errors and the root mean squared errors. The mean
squared error em is the sum of the squares of systematic error and random error,

em = eb2 + en2, (22)


1/2
and the root mean squared error er = em [8, 9, 24].
In the context of speckle pattern generation, the speckle pattern f(x, y) is a stochastic process.
Accordingly, the mean squared error em and the root mean squared error er become random
variables. The expectation of the mean squared error

E{em } = E{e2b } + E{e2n }. (23)

The systematic error can be approximated by Eq. (12), thus

E{e2b } = E{C2b } sin2 2πu0 ≈ E{Cb }2 sin2 2πu0, (24)


Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30266

where the variance of Cb is ignored for it is generally small. Following the strategy in Section
3.2, the expectation of squared random error can be approximated as

σ2 σ2
   
Var(Q) 2
E{e2n } ≈ 1+ = 1 + . (25)
E{Q} E{Q}2 M2q M2q
A combination of Eq. (23), Eq. (24), and Eq. (25) yields

σ2
 
2
E{em } ≈ E{Cb }2 sin2 2πu0 + 1 + , (26)
M2q M2q
and the corresponding root mean squared error

E{er } ≈ E{em }1/2 . (27)

Equation (26) and Eq. (27) represent estimates of the expectations of mean squared error and
root mean squared error respectively.

4. Optimal speckle radius


4.1. Pattern quality measure
A good pattern should correspond to a small overall error. Thus, a good pattern should correspond
to a small mean squared error. As the mean squared error em is a function of the subpixel position
u0 [see Eq. (26)], the integral of em over all possible positions is proposed as a pattern quality
measure [8], namely,
∫ 1
V= em (u0 ) du0 . (28)
0
In the context of speckle pattern generation, the pattern quality measure V becomes a random
variable. Substituting Eq. (26) into Eq. (28) yields

σ2
 
1 2
E{V} ≈ E{Cb }2 + 2 1 + 2 . (29)
2 M q M q
Equation (29) represents an estimate of the expectation of the pattern quality measure.

4.2. Optimal speckle radius


In practice, a speckle ψ(x, y) is generally characterized by its speckle radius R, and the speckle
density λ is generally substituted by another parameter, speckle coverage ρ = λπR2 [21]. The
speckle coverage is generally specified as 50% [26], then the issue turns to find the optimal R for
a given speckle coverage ρ.
The optimal speckle radius should minimize E{V}, namely

dE{V}
= 0. (30)
dR
A combination of Eq. (29) and Eq. (30) yields

σ2
 
dE{Cb } 4 dq
E{Cb } − 2 2 1+ 2 = 0, (31)
dR M q M q dR

where E{Cb }, given by Eq. (12), is the expectation of the amplitude of systematic error, and q,
given by Eq. (18), is the expectation of squared gradient. The optimal speckle radius Ropt should
satisfy Eq. (31).
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30267

4.3. Optimal speckle radius for Gaussian speckle patterns


Gaussian speckle patterns are widely used in practice [2, 9, 20, 37]. The profile of a Gaussian
speckle is
x2 + y2
 
ψg (x, y; R) = exp − , (32)
R2
and its spectrum is ψ̂g (ωx, ωy ; R) = πR2 exp[−R2 (ω2x + ωy2 )/4]. A Gaussian speckle is character-
ized by its radius R. The optimal radius Ropt can certainly be acquired using Eq. (31). However,
the computation may be tedious and cumbersome, and thus a simplification seems necessary.
The gradient operator d[m, n] is an approximation of the ideal gradient operator, whose transfer
function is jωx [38]; the interpolation basis function ϕ(x, y) is an approximation of the sinc
ˆ x, ωy ) is
function, whose transfer function is a rectangular function [13, 33]. Recalling that d(ω
the magnitude response of d[m, n] and ϕ̂(ωx, ωy ) is the Fourier transform of ϕ(x, y), we assume
that
ˆ x , ωy ) ≈ ω x
d(ω and ϕ̂(ωx, ωy ) ≈ 1, for − π < ωx < π, −π < ωy < π. (33)

In addition, as under-sampling should not occur,


∫ π∫ π ∫ ∞∫ ∞
ω2x ψ̂g2 (ωx, ωy ; R) dωx dωy ≈ ω2x ψ̂g2 (ωx, ωy ; R) dωx dωy = 2π 3 . (34)
−π −π −∞ −∞

With these approximations [Eq. (33) and Eq. (34)], the systematic errors [Eq. (12)] can be
approximated by
∫ π∫ π
1
E{Cb } ≈ 3 Eb (ωx, ωy )ψ̂g2 (ωx, ωy ; R) dωx dωy, (35)
2π −π −π

and the squared gradient [Eq. (18)] can be approximated by q ≈ πλ/2 = ρR−2 /2. If the correction
term is ignored, the squared random error [Eq. (25)] can be approximated by

σ2 σ2 2σ 2 R2
 
2
E{en } ≈ 2 1 + 2 ≈ 2 ≈
2
. (36)
M q M q M q ρM 2
Substituting Eq. (35) and Eq. (36) into Eq. (30) yields that the optimal radius satisfies
∫ π ∫ π
Eb (ωx, ωy )ψ̂g2 (ωx, ωy ; R) dωx dωy
−π −π
π π dψ̂g (ωx, ωy ; R) 8π 6 σ 2 R
∫ ∫
× Eb (ωx, ωy )ψ̂g (ωx, ωy ; R) dωx dωy + = 0, (37)
−π −π dR ρM 2

where Eb (ωx, ωy ) is the interpolation bias kernel, ψ̂g (ωx, ωy ; R) is the spectrum of a Gaussian
speckle, R is the speckle radius, σ = (σf2 + σg2 )1/2 is a characterization of the sensor noise, ρ is
the given speckle coverage, and M is the subset size. Equation (37) is a simplification of Eq. (31)
and is available for Gaussian speckle patterns.

5. Numerical experiments
5.1. Numerical experiments for measurement errors
In order to verify the correctness of proposed theoretical analyses, the measurement errors of
pure translation were evaluated numerically, and then a comparison between the numerical
results and the theoretical estimates was made. All source codes are uploaded to provide more
implementation details [39].
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30268

To evaluate the spatial randomness, K = 1000 Gaussian speckle patterns were generated. The
kth pattern is given by
h i2 h i2
(k) (k) 
+

N
Õ 
 x − xi y − yi


f (x, y) = ,
(k)

 

exp − 2
(38)
i=1

 R 

 
 

where N is the total number of the speckles, (xi(k), yi(k) ) is the position of the ith speckle of
the kth pattern, and R is the Gaussian speckle radius. The speckle positions satisfy a uniform
distribution over a region [−L, L] × [−L, L], where L = 50. The total number of the speckles is
N = 4π −1 ρL 2 R−2 , where ρ denotes the speckle coverage; if N is not an integer, its round is used
instead.
Then, the issue was to evaluate the systematic error eb(k) and the random error en(k) of the kth
pattern. A series of translated speckle patterns f (k) (x − u0, y) was generated; the displacement
u0 ranged from 0 to 1 pixel, with an increase of 0.05 pixels between successive patterns. The
reference pattern and the target patterns were sampled, and exact sampled values were used
to remove the quantization error. Images are inevitably contaminated by sensor noise. Sensor
noise in the reference image can be virtually removed by averaging many frames, but the
averaging approach generally cannot be applied to the target images for they must be captured
dynamically [40]. Hence, Gaussian additive white noise w[m, n] was superimposed onto the
target images exclusively; this is not fundamental because noise in the reference image does not
induce a bias error [19]. Since a typical camera signal-to-noise ratio (SNR) is 40 dB [41], the
standard deviation of the noise was specified as σ = 0.01, which is 1% of the speckle intensity. In
order to evaluate the random error, the noise addition repeated for I = 1000 times. The reference
image f (k) [m, n] and the target images g (k,i) [m, n] can be expressed as

f (k) [m, n] = f (k) (m, n), g (k,i) [m, n] = f (k) (m − u0, n) + w (k,i) [m, n], (39)

where the superscript k denotes the kth speckle pattern and i denotes the ith noise addition.
The inverse compositional Gauss-Newton (IC-GN) algorithm [35], which is the most popular
subpixel registration algorithm in the filed of DIC for its high-accuracy and high-efficiency [36,37],
was utilized to estimate the actual displacement u0 . The inverse compositional Gauss-Newton
algorithm requires gradients of reference image and subpixel intensities of target images: the
gradients were evaluated using the Barron operator [18, 36]; the intensities were interpolated
using the bicubic B-spline, which is proven far more accurate than the cubic spline [11]. The
correlation criterion chosen was SSD [see Eq. (8)]. Zero-order shape function was employed due
to the underlying deformation is rigid motion [32]. The subset was [−S, S] × [−S, S], so that the
subset size was M = 2S + 1 and there were M 2 points in the subset. The displacement u(k,i) was
evaluated, and then the measurement errors were calculated by
v
u
I h t I i2
(k) 1 Õ i
(k) 1 Õ h (k,i)
eb = u (k,i)
− u0 , en = u − u0 − eb(k) , (40)
I i=1 I − 1 i=1

where eb(k) and en(k) respectively denotes the systematic error and the random error of the kth
pattern.
Then, the spatial expectation and variance of the systematic error were given by
K K 2
1 Õ (k) 1 Õ  (k)
E{eb } = e , Var{eb } = eb − E{eb } . (41)
K k=1 b K − 1 k=1
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30269

Similarly, the spatial expectation and variance of the random error were given by
K K
2
1 Õ (k) 1 Õ  (k)
E{en } = e , Var{en } = en − E{en } . (42)
K k=1 n K − 1 k=1

The theoretical estimates of the systematic errors, the random errors, and the root mean squared
errors were evaluated using Eq. (12), Eq. (21), and Eq. (27) respectively. The double integrals
within these equations were rewritten into products of one-dimensional integrals, and then the
integrals were calculated using function integral in MATLAB.

(a ) (b )
M e a s u r e m e n t E r r o r s ( p ix e l)

0 .0 0 5

M e a s u r e m e n t E r r o r s ( p ix e l)
0 .0 0 3

0 .0 0 0 0 .0 0 0
D IC T h e o ry R = 1 .3 D IC T h e o ry R = 1 .5
e b
e b
e n
e n

-0 .0 0 5 e r -0 .0 0 3 e r

(c ) 0 .0 0 .5 1 .0 (d ) 0 .0 0 .5 1 .0
S u b p ix e l D is p la c e m e n t ( p ix e l) S u b p ix e l D is p la c e m e n t ( p ix e l)
0 .0 0 2
M e a s u r e m e n t E r r o r s ( p ix e l)

M e a s u r e m e n t E r r o r s ( p ix e l)

0 .0 0 2

0 .0 0 1
D IC T h e o ry
e b
0 .0 0 1
R = 2 .0 e n
R = 3 .0
0 .0 0 0 D IC T h e o ry e r
e b
e n 0 .0 0 0
-0 .0 0 1 e r

0 .0 0 .5 1 .0 0 .0 0 .5 1 .0
S u b p ix e l D is p la c e m e n t ( p ix e l) S u b p ix e l D is p la c e m e n t ( p ix e l)

Fig. 2. Systematic errors eb , random errors en , and root mean squared errors er of Gaussian
speckle patterns with speckle radii (a) R = 1.3 pixels, (b) R = 1.5 pixels, (c) R = 2.0 pixels,
and (d) R = 3.0 pixels. The theoretical estimates are evaluated using Eq. (12), Eq. (21), and
Eq. (27). The variability of the numerical results is due to the spatial randomness illustrated
in Fig. 1.

Systematic errors eb , random errors en , and root mean squared errors er of Gaussian speckle
patterns with speckle radii R = 1.3, 1.5, 2.0, and 3.0 pixels are shown in Fig. 2. The subset size M is
31 pixels, and the speckle coverage ρ is 50%. The theoretical estimates show excellent agreements
with the numerical results. Furthermore, it can be found that, (1) the spatial expectations of the
systematic errors are sinusoidal-shaped and decrease as the speckle radius increases; (2) the
random errors nearly remain constant and increase as the speckle radius increases; (3) the root
mean squared errors are governed by the systematic errors when the speckle radius is small
and are governed by the random errors when the speckle radius is large. These observations are
consistent with literatures and can be explained by the proposed model [11, 18, 19]. (1) Equation
(12) clearly indicates the sinusoidal-shape of systematic errors. Since the interpolation bias kernel
Eb (ωx, ωy ) is a high-pass filter [13, 14], a negative correlation exists between systematic error
and speckle radius. (2) According to Eq. (21), the random errors are independent of the actual
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30270

displacement u0 , and thus they appear as horizontal lines in Fig. 2. Equation (36) indicates that
E{en } ∼ R, and thus there is a positive correlation between random error and speckle radius. (3)
As mentioned before, an increase in speckle radius will give rise to a decrease in systematic error
and an increase in random error, leading to a bias-variance trade-off.
In following sections, the influences of subset size, speckle coverage, and speckle radius are
investigated. To characterize the systematic errors and the random errors, the actual displacement
u0 is set as 0.25 pixels.

5.1.1. Influence of subset size


Firstly, the influence of subset size on measurement errors are investigated. The subset size M
ranged from 21 to 91 pixels, in increments of 10 pixels. The speckle radius R was respectively
specified as 2 and 5 pixels; the speckle coverage ρ was respectively set as 50% and 80%.

(a 1 ) (a 2 )
0 .0 0 1 1  D IC T h e o ry  D IC T h e o ry
0 .0 0 0 2
S y s t e m a t ic e r r o r s ( p ix e l)

S y s t e m a t ic e r r o r s ( p ix e l)
5 0 % 5 0 %
0 .0 0 1 0 8 0 % 8 0 %
0 .0 0 0 1
0 .0 0 0 9
0 .0 0 0 0
0 .0 0 0 8
-0 .0 0 0 1
0 .0 0 0 7
R = 2 .0 -0 .0 0 0 2
R = 5 .0

2 1 4 1 6 1 8 1 2 1 4 1 6 1 8 1
(b 1 ) S u b s e t S iz e ( p ix e l) (b 2 ) S u b s e t S iz e ( p ix e l)
0 .0 0 8
0 .0 0 2 1  D IC T h e o ry  D IC T h e o ry
e r r o r s ( p ix e l)

e r r o r s ( p ix e l)

5 0 % 5 0 %
0 .0 0 6
8 0 % 8 0 %

0 .0 0 1 4
R = 2 .0 0 .0 0 4 R = 5 .0
R a n d o m

R a n d o m

0 .0 0 0 7 0 .0 0 2

2 1 4 1 6 1 8 1 2 1 4 1 6 1 8 1
S u b s e t S iz e ( p ix e l) S u b s e t S iz e ( p ix e l)

Fig. 3. Influence of subset size on measurement errors: systematic errors for speckle radius
(a1) R = 2.0 pixels and (a2) R = 5.0 pixels; random errors for speckle radius (b1) R = 2.0
pixels and (b2) R = 5.0 pixels. The actual displacement is u0 = 0.25 pixels. The theoretical
estimates are evaluated using Eq. (12) and Eq. (21).

Numerical results and theoretical estimates of the measurement errors corresponding to


displacement u0 = 0.25 pixels are depicted in Fig. 3. The theoretical estimates show good
agreements with the numerical results. Figure 3 exhibits that, as the subset size increases, (a) the
expectations of the systematic errors roughly remain invariant, and (b) the expectations of the
random errors will decrease. These two observations can be explained by the proposed model.
(a) Equation (12) implies that the expectation of the systematic error is specified by the gradient
estimator d[m, n], the interpolation algorithm ϕ(x, y), and the speckle profile ψ(x, y) exclusively.
As a consequence, E{Cb } is independent of subset size M and speckle coverage ρ. (b) Equation
(21) clearly indicates that increasing subset size M induces an decrease in E{en }, and the leading
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30271

order is E{en } ∼ M −1 .
The numerical results also indicate a negative correlation between the subset size and the
variability of the measurement errors. An intuitive explanation is that, with the increase of subset
size, more texture information will be exploited, thus reducing the measure uncertainty.

5.1.2. Influence of speckle coverage


Then, the influence of speckle coverage on measurement errors are investigated. The speckle
coverage ranged from 20% to 80%, in increments of 10%. The speckle radius R was respectively
specified as 2 and 5 pixels; the subset size was respectively specified as 31 and 91 pixels.

(a 1 ) (a 2 )
0 .0 0 1 2 M D IC T h e o ry M D IC T h e o ry
S y s t e m a t ic e r r o r s ( p ix e l)

S y s t e m a t ic e r r o r s ( p ix e l)
3 1 0 .0 0 0 2 3 1
9 1 9 1
0 .0 0 1 0

0 .0 0 0 0
0 .0 0 0 8

R = 2 .0 R = 5 .0
0 .0 0 0 6 -0 .0 0 0 2
2 0 4 0 6 0 8 0 2 0 4 0 6 0 8 0
(b 1 ) S p e c k le C o v e r a g e ( % ) (b 2 ) S p e c k le C o v e r a g e ( % )
0 .0 0 2 5
M D IC T h e o ry 0 .0 0 6 M D IC T h e o ry
e r r o r s ( p ix e l)

3 1 3 1
e r r o r s ( p ix e l)

0 .0 0 2 0
9 1 9 1

0 .0 0 1 5 0 .0 0 4
R a n d o m

0 .0 0 1 0
R a n d o m

R = 2 .0 0 .0 0 2
R = 5 .0
0 .0 0 0 5

2 0 4 0 6 0 8 0 2 0 4 0 6 0 8 0
S p e c k le C o v e r a g e ( % ) S p e c k le C o v e r a g e ( % )

Fig. 4. Influence of speckle coverage on measurement errors: systematic errors for speckle
radius (a1) R = 2.0 pixels and (a2) R = 5.0 pixels; random errors for speckle radius (b1)
R = 2.0 pixels and (b2) R = 5.0 pixels. The actual displacement is u0 = 0.25 pixels. The
theoretical estimates are evaluated using Eq. (12) and Eq. (21).

Numerical results and theoretical estimates of the measurement errors corresponding to


displacement u0 = 0.25 pixels are depicted in Fig. 4. The data corresponding to ρ = 20% and M
= 31 pixels is missing because it happens that there is no speckle in a subset and therefore the
iteration cannot converge. The theoretical estimates show good agreements with the numerical
results. Figure 4 exhibits that, as the speckle coverage increases, (a) the expectations of the
systematic errors nearly remain constant, and (b) the expectations of the random errors will
decrease. These two observations can be explained by the proposed model. (a) As stated in Section
5.1.1, the systematic error is independent of the speckle coverage. (b) The gradient variance q
increases with increasing speckle coverage ρ [see Eq. (18)], thus decreasing the random error
[Eq. (21)]; the leading order is E{en } ∼ ρ−1/2 .
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30272

5.1.3. Influence of speckle radius


Finally, the influence of speckle radius on measurement errors are investigated. As significant
aliasing effect occurs when R < 1.3 pixels [14], the first speckle radius was set as R = 1.3 pixels,
then R ranged from 1.5 to 5 pixels, in increments of 0.5 pixels. The subset size M was respectively
specified as 31 and 91 pixels; the speckle coverage was respectively specified as 50% and 80%.

(a 1 ) (a 2 )

S y s te m a tic e r r o r d iffe r e n c e s ( p ix e l)
0 .0 0 5 0 M = 3 1  = 5 0 %
D IC T h e o ry
S y s te m a tic e r r o r s ( p ix e l)

0 .0 0 0 6 M = 3 1  = 8 0 %
M = 3 1  = 5 0 %
M = 9 1  = 5 0 %
M = 3 1  = 8 0 %
M = 9 1  = 8 0 %
M = 9 1  = 5 0 %
0 .0 0 2 5 M = 9 1  = 8 0 % 0 .0 0 0 0

-0 .0 0 0 6

0 .0 0 0 0

(b 1 ) 1 2 3 4 5 (b 2 ) 1 2 3 4 5
S p e c k le R a d iu s ( p ix e l) S p e c k le R a d iu s ( p ix e l)
0 .0 0 1 2
0 .0 0 4
 D IC T h e o ry  D IC T h e o ry
e r r o r s ( p ix e l)

e r r o r s ( p ix e l)

5 0 % 5 0 %
8 0 % 0 .0 0 0 9 8 0 %
0 .0 0 3

M = 3 1 M = 9 1
0 .0 0 2 0 .0 0 0 6
R a n d o m

R a n d o m

0 .0 0 1 0 .0 0 0 3

1 2 3 4 5 1 2 3 4 5
S p e c k le R a d iu s ( p ix e l) S p e c k le R a d iu s ( p ix e l)

Fig. 5. Influence of speckle radius on measurement errors. (a1) Systematic errors; (a2) the
difference between systematic errors by theoretical estimations and by numerical experiments.
Random errors for subset size (b1) M = 31 pixels and (b2) M = 91 pixels. The actual
displacement u0 = 0.25 pixels. The theoretical estimates are evaluated using Eq. (12) and Eq.
(21).

Numerical results and theoretical estimates of the measurement errors corresponding to


displacement u0 = 0.25 pixels are depicted in Fig. 5. The theoretical estimates show good
agreements with the numerical results. Figure 5 indicates that, an increase in speckle radius
will give rise to (a) a decrease in systematic error and (b) an increase in random error. Theses
two observations can be explained by the proposed model. (a) It follows from Eq. (35) that the
systematic error is determined by the integral of the product of the interpolation bias kernel
Eb (ωx, ωy ) and the power spectrum | ψ̂g (ωx, ωy ; R)| 2 . With the increase of speckle radius, more
energy will concentrate on the low-frequency domain. Since the interpolation bias kernel is a
high-pass filter [13, 14], the systematic error will decrease. (b) It follows from Eq. (36) that the
random error √

E{en } ≈ √ R, (43)
ρM
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30273

thus the random


√ error is approximately proportional to the speckle radius R with a proportionality
constant 2σM −1 ρ−1/2 . The random errors in Fig. 5(b1)-(b2) are normalized by a factor
Mσ −1 (ρ/2)1/2 and then are illustrated in Fig. 6. Figure 6 implies that the approximation [Eq.
(43)] is more accurate when the subset size is large. This is because the linearity is based on an
elimination of the correction term, which is related to the variance, and meanwhile the variability
of a small subset size is more significant.

M = 3 1  = 5 0 %
6 M = 3 1  = 8 0 %
M = 9 1  = 5 0 %
( p ix e l) 5 M = 9 1  = 8 0 %
2 
e ≈ R
n

n
4 M
e


1 /2
M σ- 1 ( ρ/ 2 )

 M
⇒ e n ≈R
3 2 

y = R
2

1
1 2 3 4 5
S p e c k le R a d iu s ( p ix e l)

Fig. 6. The normalized random errors are approximately proportional to the speckle radius.

5.2. Numerical experiments for optimal speckle radius


In order to validate the formulas of pattern quality measure V [Eq. (29)] and optimal speckle
radius Ropt [Eq. (31) and Eq. (37)], numerical experiments were conducted and then a comparison
between the numerical results and the theoretical estimates was made.
The numerical method utilized is similar to that in Section 5.1; the difference is that, in order
to enhance the evaluation accuracy of the pattern quality measure V, 100 translated patterns were
generated, with a displacement of 0.01 pixels between successive patterns. Recall that in Section
5.1, there are 20 translated patterns and the relative displacement is 0.05 pixels. The mean square
errors were evaluated and then the pattern quality measure V, an average of the mean squared
errors, was numerically evaluated by the trapezoidal rule [42]. Theoretical estimations of the
pattern quality measure are based on Eq. (29). The optimal speckle radii Ropt are numerically
evaluated using Eq. (31) and Eq. (37) respectively by the bisection method [42].
The pattern quality measure by numerical experiments and by theoretical estimations are
illustrated in Fig. 7. The subset size M was specified as 31 and 91 pixels respectively; the speckle
coverage ρ was specified as 50% and 80% respectively; the speckle radius R was firstly selected as
1.3 pixels, and then R ranged from 1.5 to 5 pixels, with an increase of 0.25 pixels. The numerical
results show good agreements with the theoretical estimations. Evidently, there is a trade-off
between systematic errors and random errors: when the speckle radius is excessively small, the
systematic error is significant; when the speckle radius is too large, the random error is significant.
Therefore, an optimal speckle radius exists, which can be estimated using Eq. (31), the full form,
or Eq. (37), the approximation form for Gaussian speckle patterns. For the cases in Fig. 7, the
optimal radii given by Eq. (31) are 1.977, 2.075, 2.453, and 2.570 pixels respectively, and the
optimal radii given by Eq. (37) are 1.974, 2.071, 2.448, and 2.564 pixels respectively, which
are almost the same as Eq. (31). The corresponding speckle sizes are 4 ∼ 5 pixels, which are
consistent with the results in other literatures [1, 20, 22]. It can also be found that, as the increase
of subset size or speckle coverage, the optimal speckle radius will increase. This is because
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30274

-5
(a ) 2 .0
x 1 0 -5
( b ) 1 .5 x 1 0

P a tte r n q u a lity m e a s u r e ( p ix e l2 )

P a tte r n q u a lity m e a s u r e ( p ix e l2 )
D IC D IC
T h e o ry T h e o ry
1 .5
1 .0
R o p t = 1 .9 7 7 R o p t = 2 .0 7 5
1 .0

0 .5
0 .5
M = 3 1 M = 3 1
 = 5 0 %  = 8 0 %
0 .0 0 .0
1 2 3 4 5 1 2 3 4 5
(c ) x 1 0 -5 S p e c k le R a d iu s ( p ix e l) (d ) x 1 0 -5 S p e c k le R a d iu s ( p ix e l)
P a tte r n q u a lity m e a s u r e ( p ix e l2 )

P a tte r n q u a lity m e a s u r e ( p ix e l2 )
D IC D IC
1 .0 T h e o ry 1 .0 T h e o ry

R o p t = 2 .4 5 3 R o p t = 2 .5 7 0
0 .5 0 .5
M = 9 1 M = 9 1
 = 5 0 %  = 8 0 %

0 .0 0 .0
1 2 3 4 5 1 2 3 4 5
S p e c k le R a d iu s ( p ix e l) S p e c k le R a d iu s ( p ix e l)

Fig. 7. Relationship between pattern quality measure and speckle radius: (a) subset size
M = 31 pixels, speckle coverage 50%; (b) subset size M = 31 pixels, speckle coverage 80%;
(c) subset size M = 91 pixels, speckle coverage 50%; (d) subset size M = 91 pixels, speckle
coverage 80%. The theoretical estimations of the pattern quality measure are evaluated using
Eq. (29) and the optimal speckle radii Ropt are evaluated using Eq. (31).

the systematic errors are independent of subset size and speckle coverage, while the random
error will decrease if the subset size or the speckle coverage increase. Besides, it seems that the
variance of V will increase in response to the increase of E{V}.
As a whole, the numerical results demonstrate the correctness of proposed theoretical model.

6. Discussion
6.1. Advantages
The theoretical model is superior to the traditional numerical methods in three aspects. (1)
Generality: the theoretical model is applicable to all parameters; by contrast, conducting a
simulation for every possible parameter is impossible. (2) Efficiency: it takes days to evaluate
the numerical results in Fig. 7; by contrast, the theoretical estimates can be attained in about 1
second. (3) Deep understanding: the model explains why the systematic error is independent of
the speckle density and why the random error is approximately proportional to the speckle radius.

6.2. Limits and outlook


Although the proposed theoretical model has the advantages as described above, there are still four
limits exist. (1) The proposed model fails to characterize the spatial variances of measurement
error and pattern quality measure because of the difficulties faced in mathematics. (2) The optimal
speckle radius provided by the model is optimal in a sense of statistics. How to optimize the
Vol. 25, No. 24 | 27 Nov 2017 | OPTICS EXPRESS 30275

position of every speckle is still a problem. (3) The proposed model cannot suggest an optimal
speckle coverage due to the assumption that the noise σ is a constant. In practice, the image
noise depends on the intensity [16, 41]. Thus, a more appropriate assumption may be that the
signal-to-noise ratio is a constant. However, it will make the theoretical analyses and the numerical
simulations far more complicated. (4) The proposed model can suggest an optimal speckle radius
for pure translation. Nevertheless, whether this value can be instructive in cases of deformations
or rotations is still unclear. These problems need further research.
The key to our model is to derive the spatial expectations of systematic errors and random
errors. For the systematic errors, the key is to derive the pattern power spectrum, which is the
Fourier transform of the auto-correlation function due to Wiener-Khinchin theorem. For the
random errors, the key is to derive the expectation of squared gradient. Speckle patterns which are
composed of random positioned speckles are studied because of their wide application. However,
there is a great hope that this method can be applied to other types of pattern, provided that the
power spectrum and the expectation of squared gradient are known.

7. Conclusion
This paper concentrates on speckle patterns which are composed of random positioned speckles.
A statistical model for speckle pattern optimization is built and validated. All source codes are
uploaded and ready to be downloaded. The conclusions of this work are drawn as follows:
(1) The process of speckle pattern generation is essentially a filtered Poisson process.
(2) Spatial expectations of systematic errors [Eq. (12)], random errors [Eq. (21)], and overall
errors [Eq. (26) and Eq. (27)] are characterized mathematically. The expectations of systematic
errors are independent of speckle coverage and subset size; the expectations of random errors
are approximately proportional to the speckle radius. The formulas proposed will facilitate the
estimation of metrological performance.
(3) The relationship between spatial expectation of pattern quality measure and speckle pattern
generation parameters is deduced [Eq. (29)]. Assuming the noise variance remains constant, a
formula for the optimal speckle radius is derived [Eq. (31)] and a simplified version is presented
for Gaussian speckle patterns [Eq. (37)].
The original contribution of this work is the presentation of a theoretical model for speckle
pattern optimization. As the mathematical derivations merely needs pattern power spectrum and
expectation of squared gradient, there is a great hope that this method can be applied to patterns
of other type. The primary motivation of this research is to establish a speckle pattern generation
standard for DIC. However, as the model is general, the authors believe that scholars in other
optical measurement communities, such as PIV and speckle metrology, will benefit from these
discussions.

Funding
National Natural Science Foundation of China (11702287, 11332010, 11627803, 11472266);
Strategic Priority Research Program of the Chinese Academy of Sciences (XDB22040502).

You might also like