SoftwareX 11 (2020) 100391

Contents lists available at ScienceDirect

SoftwareX
journal homepage: www.elsevier.com/locate/softx

Original software publication

µDIC: An open-source toolkit for digital image correlation

∗
Sindre Nordmark Olufsen a , , Marius Endre Andersen b , Egil Fagerholt a
a
Structural Impact Laboratory (SIMLab), Department of Structural Engineering, NTNU, Norwegian University of Science and
Technology, NO-7491 Trondheim, Norway
b
SINTEF Industry, Department of Materials and Nanotechnology, NO-7034 Trondheim, Norway

article info a b s t r a c t

Article history: We here present a Digital Image Correlation toolkit, formulated as a Python package. This package
Received 29 May 2019 aims at providing a complete toolkit for performing DIC analysis on experimental data, performing
Received in revised form 25 October 2019 virtual experiments, as well as a framework for further development. A suite of tools for generating
Accepted 17 December 2019
synthetic speckle images, modelling of sensor artefacts and deformation of images by displacement
Keywords: fields, are included. The virtual experiments are used as a part of the accuracy assessment of the toolkit
Digital Image Correlation as well as for testing during development. B-spline elements are employed for the discretisation of
B-splines the displacement fields and allow the polynomial order and degree of continuity to be controlled by
Speckle image generation the user.
Virtual experiments © 2019 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license
Python (http://creativecommons.org/licenses/by/4.0/).

Code metadata

Current code version 0.1.1

Permanent link to code/repository used for this code version https://github.com/ElsevierSoftwareX/SOFTX_2019_193
Legal Code License MIT
Code versioning system used GIT
Software code languages, tools, and services used Python, Numpy, Scipy, Matplotlib, Numba
Compilation requirements, operating environments & Cross platform
If available Link to developer documentation/manual https://mudic.readthedocs.io/en/latest/
Support email for questions [email protected]

1. Motivation and significance DIC has been a valuable tool in a vast range of scientific fields
such as characterisation of biological materials [3–6], detection of
Digital Image Correlation (DIC) has become a popular tool tumours [7], nanoscale investigation of materials using TEM [8],
for quantification of surface deformations, due to its ease of use characterisation of mechanical properties of polymers [9–11], and
and ability to provide local deformation measurements even for many others.
inhomogeneous deformation fields [1,2]. DIC is performed by pro- There are several variants of DIC, the most common be-
cessing a series of images captured during the deformation of the ing subsets-based DIC (local DIC) and Finite-Element-based DIC
object of interest, and a mapping between the image coordinates (global DIC). Subset-based DIC attempts to correlate independent
of an image of the undeformed (reference) object and the image subsets of the image, often being in the order of ten to thirty
coordinates of an image of the deformed object is calculated. This pixels in each direction. In order to cover a large portion of
operation is often facilitated by applying a speckle-like pattern on the specimen surface, a large number of subsets are used. FE-
the surface of the specimen. In order to obtain such a mapping, based DIC, on the other hand, discretises the deformation field
an optimisation scheme is employed, searching for the mapping using finite elements, approximating the inter-nodal coordinates
which gives the best correlation between a reference image and by interpolation. FE-based DIC normally enforces displacement
the current image. The coordinate mapping is in turn is used to continuity, but by using higher order element formulations [12,
calculate the strain fields. 13], B-splines [14] or NURBS [15], higher order continuity can
be enforced. Interestingly, a performance comparison of subset-
∗ Corresponding author. and FE-based DIC was done by Hild et al. [16] and later by Pan
E-mail address: [email protected] (S.N. Olufsen). et al. [17], reaching opposing conclusions. Recent advances in the

https://doi.org/10.1016/j.softx.2019.100391
2352-7110/© 2019 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
2 S.N. Olufsen, M.E. Andersen and E. Fagerholt / SoftwareX 11 (2020) 100391

field of DIC include self-adaptive element order [18] and more 2.1. Software architecture
computationally efficient formulations [19].
Commercial DIC software such as ISI-Sys VIC and GOM- The µDIC tookit is formulated as a set of packages:
correlate, and open source DIC codes such as DICE [20] and • IO: Tools for image import and export
nCorr [21] are now available. For scientific use, it can be appealing • VirtualLab: Tools for synthetic data generation
to use open source software, as insight into the internal architec- • Mesher: Tools for mesh generation and manipulation
ture of the software is available. However, at the time of writing, • Correlator: The digital image correlation routines
the available open source DIC codes were either formulated in • PostProcessor: Tools for post-processing and visualisation
C++, calling for significant programming knowledge, bound to
commercial software such as MATLAB, or formulated as scripts The packages can be combined, providing a complete process-
with very limited testing and documentation. ing pipeline tailored to the user’s needs. A typical pipeline and
This project aims to provide a complete and easy to use toolkit the most relevant package components are shown in Fig. 1 and
run as follows:
for 2D digital image correlation, that is verified and provided with
An ImageStack object is generated either by loading a set
automated testing, implemented in an accessible programming
of images by using imagestack_from_folder or by constructing a
language. Python was chosen for this task, aided by NumPy [22],
VirtualExperiment. A Mesher is instantiated and a mesh object is
SciPy [23] and Numba [24] for fast array operations. Visualisation obtained by passing the ImageStack to the mesh method. The re-
is provided by Matplotlib [25]. sulting Mesh object and ImageStack is then stored in the DICInput
Motivated by the improved accuracy and flexibility reported object. The DICAnalysis factory verifies DICInput and a DICJob is
by Cheng et al. [14], B-splines are used to discretise the defor- constructed. The DIC analysis can now be initiated by calling
mation field. B-splines allows for a very compact and flexible the run method. The resulting DICRes object contains all relevant
formulation where the user can specify the degree of the in- output from the DIC analysis, and field variables can be calculated
terpolation polynomials and the degree of continuity as most by DICPost class and visualised by the plot function.
appropriate for their application. Linear and parabolic element
formulations such as Q4 and Q8 are subsets of the B-splines and 2.2. Software functionality
are therefore an integrated part of the package.
The core functionality of the main packages in the toolkit is
A modified Newton–Raphson iterative solver [12] is used to
presented in the following.
minimise the sum of squared differences between the reference
and current image. During the optimisation process, the inter- 2.2.1. IO
pixel values of the image are obtained by interpolation using The IO package provides all the tools needed for importing
bi-cubic or bi-quintic splines. images and saving objects. The most typical tasks, such as making
An additional but important component is a toolkit for gener- an image stack of all images in a folder with a given extension,
ating and deforming speckle images, including camera artefacts are made easier by providing factories such as imagestack_from_
such as fill-factor and noise. This toolkit is based on the ideas folder.
presented by Orteu et al. [26]. The toolkit allows for assessment
of element formulations, aliasing artefacts and filtering, and is 2.2.2. Virtual lab
an integral part of the provided suite of integration tests. Addi- The VirtualLab package provides tools for the generation of
tional validation is provided by experimental data where other synthetic speckle images, deformation of speckle images using
prescribed deformation fields and modelling of camera artefacts
techniques have been used to measure deformations.
such as fill-factor and pixel eccentricity. The virtual lab toolkit
The user can implement this package in the daily workflow
is based on the work of Orteu et al. [26]. The main modules of
by adopting one of the provided example-scripts, where in most
the package and their intended usage are illustrated in Fig. 2 and
cases the only required alteration is the path to the folder con- described below:
taining the images to be analysed.
This toolkit was conceived as a Matlab-code during the PhD • The Speckle module contains tools for generating spray-
study of Dr Andersen [12] and has been used in several stud- paint-like speckles using a Perlin noise [29] based algorithm
ies [10–12]. It was later rewritten in Python by the corresponding or the Rosta algorithm found in Appendix. Tools for gener-
author and has been applied in [9,27]. ating speckles formed by circular dots or bi-harmonic fields
are also included. The resolution, speckle size and degree of
smoothing can be controlled by the user.
2. Software description • The Deformer module contains tools for distorting the
speckle images according to a prescribed displacement field
The µDIC toolkit provides a collection of high-level functions or by a deformation gradient. Inversion of the displacement
and classes, used for importing images, generating synthetic input function is done by employing a Newton solver. Bi-quintic
data, meshing, image correlation and post-processing. The API is splines are used to interpolate the grey-scale values of the
constructed such that the functions and classes can be used as image.
stand-alone components. Factories are included in all packages, • The Downsampler module contains tools for down-sampling
of images, including sensor artefacts such as fill-factor and
providing default behaviour for the most common tasks. The
pixel eccentricity. The sensor is modelled as a grid where
toolkit can be used in the daily workflow when performing real
each pixel of the sensor covers a finite region of the image.
experiments or used for exploring the impact of different factors
The grey scale values of the image covered by a pixel of
such as element formulations, speckle design and filtering by the sensor are then integrated, giving the grey scale value
performing virtual experiments. An interactive experience can be of the sensor pixel. Each pixel on the sensor can be offset
obtained by using IPython [28] or Jupyter notebooks. Testing is an by a random value picked from a Gaussian distribution
integral part of this toolkit, and virtual experiments, simulating with a prescribed standard deviation. The position of the
know deformation fields, are used as integration tests along with super-sampled points covered by each pixel of the sensor
other unit tests. is calculated, as shown in [26].
 S.N. Olufsen, M.E. Andersen and E. Fagerholt / SoftwareX 11 (2020) 100391 3

Fig. 1. An example pipeline showing the main packages and the most relevant components.

Fig. 2. Typical use of VirtualLab to create speckle images with a given deformation and noise level. Module names and constructors are shown.

• The NoiseInjector module is used to introduce noise to the correlation routine follows [31] except for the chosen correlation
downsampled images. The user can define a noise model criterion.
and the standard deviation of the noise distribution. An The objective of the correlation routine is to find the current
additive Gaussian noise model is provided with the toolkit, image coordinates for each pixel x such that:
and the user can easily add other models.
IC (x) = IR (x0 ) (3)
2.2.3. Mesher
where IC (x) is the grey-scale intensities of the current image at
The Mesher package contains the tools for mesh generation
the pixel coordinates x and IR (x0 ) is the grey-scale intensities of
and manipulation. B-splines are used to discretise the deforma-
the reference image at the initial pixel coordinates x0 . In real
tion field into a bi-directional grid of k x l control points. The
formulation and algorithms are taken from [30]. The coordinates world conditions, Eq. (3) cannot be satisfied directly, and it is
x(u, v ) on the B-spline surface is given by: therefore approximated by using the sum of squared differences
(SSD) between the grey scale values of the current and reference
k l
∑ ∑ image as correlation criterion. By using Eq. (2), the SSD correlation
x(u, v ) = Ni,p (u)Nj,q (v )mij (1) criterion can be written in terms of control point positions as:
i=0 j=0
R = ∥I¯C (n) − I¯R (n0 )∥2 (4)
where Ni,p and Nj,q are the B-spline basis functions, u and v are
the spline coordinates in the range [0, 1] and p and q denotes where n0 denotes the initial control point position. The approxi-
the polynomial order. mij contains the coordinates of the control mate solution of Eq. (3) is then found as:
points along the bi-directional grid axes of the spline surface. By
rearranging, Eq. (1) can be written on the form: min R (5)
n

x(u, v ) = A(u, v )n (2) In order to find the control point positions n which satis-
where n is a vector contains all the control point positions and fies Eq. (5), a Newton Raphson scheme is used:
A(u, v ) is a matrix containing the values of the B-spline basis ∇∇ R(ni )[ni+1 − ni ] = −∇ R(ni ) (6)
functions.
where i denotes the increment. The term ∇∇ R(ni ) is simplified by
2.2.4. Correlator assuming that I¯C (n) ≈ I¯R (n0 ) as proposed by [31] and by using the
The Correlator package contains the Digital Image Correlation grey scale gradients of the reference image instead of the current
routines. A modified Newton scheme is used to minimise the image, giving:
sum of squared differences between the grey scale values of the
reference image and the current image. The derivation of the ni+1 = ni + (bT b)−1 bT ∆Ī¯ i (7)
4 S.N. Olufsen, M.E. Andersen and E. Fagerholt / SoftwareX 11 (2020) 100391

where 3.2. Example 2: Performing a virtual experiment

∂ IR
b= A (8) When the user wants to evaluate the performance of the
∂x
digital image correlation routine or assess the influence of pa-
Here, ∆Ī i is the difference in grey-scale value between the current rameters such as speckle design and element formulations, the
and reference frame. VirtualLab module can be used. We here show an example where
The inter-pixel values of ĪC (ni ) and ĪR (n0 ) are determined by a speckle with a resolution of 16 Mpx is generated using the
means of bi-cubic or bi-quintic spline interpolation, which has Rosta algorithm and deformed by a bi-harmonic displacement
been demonstrated to minimise the bias error [32]. For details field. After deformation, the speckle is down-sampled by a factor
and discussion regarding the derivation of the correlation routine, of 8 and noise is added. DIC is then performed on the generated
see [31] and [12]. data set. The code for performing this is shown below:

2.2.5. Post processing

Based on the B-spline surface xt (u, v ) at frame t, the deforma-
tion gradient for each material point within the surface can be
calculated as [33]:
⏐ ⏐⏐ ⏐−1
⏐x xt ,v ⏐⏐ ⏐⏐x0,u y0,v ⏐⏐
Ft = ⏐⏐ t ,u (9)
yt ,u yt ,v ⏐ ⏐x0,u y0,v ⏐

where the entries of Ft are partial derivatives of the spatial

coordinate component fields with respect to u and v . Based on the
resulting deformation gradient Ft , the right Cauchy–Green tensor
is calculated:

Ct = FTt Ft = U2t (10)

Ut is found by calculating the square root of the eigenvalues of U2t

as found by spectral decomposition. The Hencky strain tensor εt
(often referred to as true or logarithmic strain) is then determined
as:

εt = ln(Ut ) (11)

Note that the eigenvectors of εt and U2t

are now coaxial. The
Hencky strain tensor εi is then rotated into the coordinate frame
of the image, giving the Hencky strain components in image
coordinates.

3. Illustrative examples

3.1. Example 1: Processing experimental data

Processing of experimental data is one of the routine tasks

of a scientist. We here show an example where Digital Image
Correlation is performed on images that were acquired during a
tensile test of a flat dog-bone specimen made of Docol 600L. The The x-component of the imposed displacement field and the
output of the procedure is the longitudinal true strain field shown deviation between the DIC results and the imposed field is shown
in Fig. 3. in Fig. 4, demonstrating how the accuracy of the DIC routines can
The complete script used to produce these results is shown be assessed.
below:
4. Impact

Digital Image Correlation has been widely adopted in several

fields of science due to its practicality in real-world experimental
setups as well as the amount of detail which can be obtained
from the measurements. The current tool aims to provide an
easy-to-use, transparent and powerful toolkit for researchers and
industry. The toolkit can be applied to a broad range of research
problems where quantification of surface deformations is of in-
terest. The popularity of Python makes it an ideal language for
such a tool, motivating adoption, modification and extension. By
distributing the toolkit via the Python package manager along
with the unit test, and providing the user with examples and
documentation, the threshold for taking the toolkit into use is
lowered. The authors would like to highlight the following:
 S.N. Olufsen, M.E. Andersen and E. Fagerholt / SoftwareX 11 (2020) 100391 5

• Evaluation of the internals such as the correlation routines,

element formulations and interpolation routines, is here
facilitated through the use of the VirtualLab module.
• Due to the use of the VirtualLab module in the integration
tests, modification and extension of the toolkit is made
robust, as the impact of the modification is easily evaluated.
• As the toolkit is made available as a Python package, all data
processing can be done without changing to other systems.
This motivates the adoption of the vast ecosystem of Python
packages for tasks such as visualisation, optimisation and
machine learning.
• Large sets of experimental data are easily processed by
scripting, and example scripts for this task are provided
along with the package.
• As the toolkit is distributed under the MIT license, commer-
cial adoption of the package and derivatives of the package
are facilitated.

5. Conclusions
Fig. 3. Results from the digital image correlation pipeline showing the longitu-
We here present a toolkit for Digital Image Correlation and
dinal true strain for a deformed Docol 600L dog-bone specimen at three stages
of deformation. outline its capabilities. The toolkit includes tools for synthetic
input generation, a key feature for validation and further devel-
opment. B-spline discretisation of the deformation field is used,
providing a flexible framework where the degree of the polyno-
mials and the order continuity can be specified. Python is chosen
for the implementation of the toolkit, making it accessible to a
broader user group and motivating use and modification.

Declaration of competing interest

The authors declare that they have no known competing finan-

cial interests or personal relationships that could have appeared
to influence the work reported in this paper.

Acknowledgements

The authors gratefully appreciate the financial support from

the Research Council of Norway through the Centre for Advanced
Structural Analysis, Project No. 237885 (SFI-CASA).

Appendix. The Rosta algorithm

The following thought experiment formed this algorithm: If a

haze of droplets from a spray nozzle travel through unsteady air
and land on a surface, they may consolidate into larger droplets.
The probability for a droplet to consolidate with another droplet
should be coupled to the density of droplets in close proximity. If
the density of droplets is higher than some threshold value, the
droplets are anticipated to interact and to form a cluster. Based
on this thought experiment, the following algorithm is proposed:
A comparison of a subsection of a speckle generated by the
algorithm and a real speckle is shown in Fig. A.5.

Fig. 4. Results from the virtual experiment. The x-component of the imposed
displacement field is shown in figure (a) and the deviation between the DIC
results and the imposed field is shown in (b).

• B-splines discretisation of the deformation field provides a

flexible formulation where the user can control the degree
of the polynomials and the order of continuity. Fig. A.5. A comparison of (a) a synthetic speckle and (b) a real speckle.
6 S.N. Olufsen, M.E. Andersen and E. Fagerholt / SoftwareX 11 (2020) 100391

Algorithm 1: The Rosta algorithm [11] Johnsen Joakim, Grytten Frode, Hopperstad Odd Sture, Clausen Arild Holm.
Experimental set-up for determination of the large-strain tensile behaviour
Input: s: an integer representing the image size; of polymers at low temperatures. Polym Test 2016;53:305–13.
σp : the size of the proximity kernel; [12] Andersen Marius. An Experimental and Numerical Study of Thermoplastics
σs : the size of the smoothing kernel; at Large Deformations (PhD thesis), 2016.
[13] Ma Shaopeng, Zhao Zilong, Wang Xian. Mesh-based digital image correla-
n: the number of repetitions t: fill fraction tion method using higher order isoparametric elements. J Strain Anal Eng
Output: I: speckle image Des 2012;47(3):163–75.
make an empty image, I with size s x s ; [14] Cheng Peng, Sutton Michael A, Schreier Hubert W, McNeill Stephen R. Full-
make gaussian kernel kp with σ = σp ∗ s; field speckle pattern image correlation with B-spline deformation function.
Exp Mech 2002;42(3):344–52.
make gaussian kernel ks with σ = σs ∗ s; [15] Réthoré J, Elguedj T, Simon P, Coret M. On the use of NURBS functions for
for n layers do displacement derivatives measurement by digital image correlation. Exp
make random number field, f of size s; Mech 2010;50(7):1099–116.
calculate g = f ∗ kp ; [16] Hild F, Roux S. Comparison of local and global approaches to digital image
correlation. Exp Mech 2012;52(9):1503–19.
scale g to unity ; [17] Wang Bo, Pan Bing. Subset-based local vs. finite element-based global
find v such that a fraction of t pixels are found by g < v digital image correlation: A comparison study. Theor Appl Mech Lett
move values of g < v into I; 2016;6(5):200–8.
end [18] Wittevrongel L, Lava P, Lomov SV, Debruyne D. A self adaptive global
digital image correlation algorithm. Exp Mech 2015;55(2):361–78.
calculate I = I ∗ ks ; [19] Pan B, Li K, Tong W. Fast, robust and accurate digital image
correlation calculation without redundant computations. Exp Mech
2013;53(7):1277–89.
[20] Turner Dan, Crozier Paul, Reu Phil. Digital image correlation engine, version
References 00. 2015.
[21] Blaber J, Adair B, Antoniou A. Ncorr : Open-source 2d digital image
correlation matlab software. 2015.
[1] Sutton Michael A, Orteu Jean-José, Schreier Hubert W. Image Correlation
[22] Oliphant Travis E. Guide to NumPy. 2nd ed.. USA: CreateSpace Independent
for Shape, Motion and Deformation Measurements, Vol. 53. Springer
Publishing Platform; 2015.
Science+Business Media; 2009, p. 1689–99, 9.
[23] Jones Eric, Oliphant Travis, Peterson Pearu, et al. SciPy: Open source
[2] Pan Bing, Qian Kemao, Xie Huimin, Asundi Anand. Two-dimensional digital
scientific tools for Python.
image correlation for in-plane displacement and strain measurement: A
[24] Lam Siu Kwan, Pitrou Antoine, Seibert Stanley. Numba: A LLVM-based
review. Meas Sci Technol 2009;20(6).
python JIT compiler. In: Proceedings of the Second Workshop on the LLVM
[3] Zhang Dongsheng, Arola Dwayne D. Applications of digital image
Compiler Infrastructure in HPC - LLVM ’15, 2015, p. 1–6.
correlation to biological tissues. J Biomed Opt 2004;9(4):691.
[25] Hunter JD. Matplotlib: A 2D graphics environment. Comput Sci Eng
[4] Rizzuto E, Carosio S, Del Prete Z. Characterization of a digital image
2007;9(3):90–5.
correlation system for dynamic strain measurements of small biological
[26] Orteu Jean-José, Garcia Dorian, Robert Laurent, Bugarin Florian. A speckle
tissues. Exp Tech 2016;40(2):743–53.
texture image generator. In: Speckle06: Speckles, From Grains to Flowers,
[5] Rizzuto Emanuele, Carosio Silvia, Musaro Antonio, Del Prete Zaccaria. A
Vol. 6341. 2006, p. 63410H, September 2006.
Digital Image Correlation based technique to control the development of a
[27] Olufsen Sindre Nordmark, Clausen Arild Holm, Breiby Dag Werner, Hop-
skeletal muscle engineered tissue by measuring its surface strain field.
perstad Odd Sture. X-ray computed tomography investigation of dilation
In: 2015 IEEE International Symposium on Medical Measurements and
of mineral-filled PVC under monotonic loading. Mechanics of Materials
Applications, MeMeA 2015 - Proceedings. 2015, p. 314–8.
2020;142:103296.
[6] Palanca Marco, Tozzi Gianluca, Cristofolini Luca. The use of digital
[28] Pérez Fernando, Granger Brian E. IPython: A system for interactive sci-
image correlation in the biomechanical area: A review. Int Biomech
entific computing python: An open and general- purpose environment.
2016;3(1):1–21.
Comput Sci Eng 2007;9(3):21–9.
[7] Krehbiel JD, Lambros J, Viator JA, Sottos NR. Digital image correlation for
[29] Perlin Ken. Improving noise. ACM Trans Graph 2002;21(3):681–2.
improved detection of basal cell carcinoma. Exp Mech 2010;50(6):813–24.
[30] Piegl Les, Tiller Wayne. The NURBS Book. Monographs in Visual
[8] Liu Yang, Xia Shuman, Mao Scott X, Pan Zhipeng, Wang Xueju, Zhu Ting,
Communication, Springer Series; 1997.
Wang Jiangwei, Fan Feifei. Nanoscale deformation analysis with high-
[31] Vend Roux G, Knauss WG. Submicron deformation field measurements:
resolution transmission electron microscopy and digital image correlation.
Part 2. Improved digital image correlation. Exp Mech 1998;38(2):86–92.
J Appl Mech 2015;82(12):121001.
[32] Sutton Michael A, Schreier Hubert W, Braasch Joachim R. Systematic errors
[9] Olufsen Sindre, Clausen Arild Holm, Hopperstad Odd Sture. Influence of
in digital image correlation caused by intensity interpolation. Opt Eng
stress triaxiality and strain rate on stress-strain behaviour and dilation of
2000;39(November 2000):2915–21.
mineral-filled PVC. Polym Test 2019;75:350–7.
[33] Tustison Nicholas J, Amini Amir A. Analysis Of 4-D Cardiac Mr Data
[10] Johnsen Joakim, Grytten Frode, Hopperstad Odd Sture, Clausen Arild Holm.
With Nurbs Deformable Models: Temporal Fitting Strategy And Nonrigid
Influence of strain rate and temperature on the mechanical behaviour of
Registration. New York: Springer; 2007, p. 493–534.
rubber-modified polypropylene and cross-linked polyethylene. Mech Mater
2017;114:40–56.