Lecture Notes in Networks and Systems 144 Tawfik Masrour Ibtissam El Hassani Anass Cherrafi Editors Artificial Intelligence and Industrial Applications Artificial Intelligence Techniques for Cyber-Physical, Digital Twin Systems and Engineering Applications
30
Embed
Tawfik Masrour Ibtissam El Hassani Anass Cherrafi Editors ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Lecture Notes in Networks and Systems 144
Tawfik MasrourIbtissam El HassaniAnass Cherrafi Editors
Artificial Intelligence and Industrial ApplicationsArtificial Intelligence Techniques for Cyber-Physical, Digital Twin Systems and Engineering Applications
Lecture Notes in Networks and Systems
Volume 144
Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,Warsaw, Poland
Advisory Editors
Fernando Gomide, Department of Computer Engineering and Automation—DCA,School of Electrical and Computer Engineering—FEEC, University of Campinas—UNICAMP, São Paulo, Brazil
Okyay Kaynak, Department of Electrical and Electronic EngineeringBogazici University, Istanbul, Turkey
Derong Liu, Department of Electrical and Computer Engineering, Universityof Illinois at Chicago, Chicago, USA, Institute of Automation, Chinese Academyof Sciences, Beijing, China
Witold Pedrycz, Department of Electrical and Computer Engineering,University of Alberta, Alberta, Canada, Systems Research Institute,Polish Academy of Sciences, Warsaw, Poland
Marios M. Polycarpou, Department of Electrical and Computer Engineering,KIOS Research Center for Intelligent Systems and Networks, University of Cyprus,Nicosia, Cyprus
Imre J. Rudas, Óbuda University, Budapest, Hungary
Jun Wang, Department of Computer Science, City University of Hong Kong,Kowloon, Hong Kong
The series “Lecture Notes in Networks and Systems” publishes the latestdevelopments in Networks and Systems—quickly, informally and with high quality.Original research reported in proceedings and post-proceedings represents the coreof LNNS.
Volumes published in LNNS embrace all aspects and subfields of, as well as newchallenges in, Networks and Systems.
The series contains proceedings and edited volumes in systems and networks,spanning the areas of Cyber-Physical Systems, Autonomous Systems, SensorNetworks, Control Systems, Energy Systems, Automotive Systems, BiologicalSystems, Vehicular Networking and Connected Vehicles, Aerospace Systems,Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power Systems,Robotics, Social Systems, Economic Systems and other. Of particular value to boththe contributors and the readership are the short publication timeframe and theworld-wide distribution and exposure which enable both a wide and rapiddissemination of research output.
The series covers the theory, applications, and perspectives on the state of the artand future developments relevant to systems and networks, decision making, control,complex processes and related areas, as embedded in the fields of interdisciplinaryand applied sciences, engineering, computer science, physics, economics, social, andlife sciences, as well as the paradigms and methodologies behind them.
** Indexing: The books of this series are submitted to ISI Proceedings,SCOPUS, Google Scholar and Springerlink **
More information about this series at http://www.springer.com/series/15179
EditorsTawfik MasrourDepartment of Mathematicsand Computer ScienceNational Graduate Schoolfor Arts and CraftsMeknes, Morocco
Anass CherrafiDepartment of Industrialand Manufacturing EngineeringNational Graduate Schoolfor Arts and CraftsMeknes, Morocco
Ibtissam El HassaniDepartment of Industrialand Manufacturing EngineeringNational Graduate Schoolfor Arts and CraftsMeknes, Morocco
ISSN 2367-3370 ISSN 2367-3389 (electronic)Lecture Notes in Networks and SystemsISBN 978-3-030-53969-6 ISBN 978-3-030-53970-2 (eBook)https://doi.org/10.1007/978-3-030-53970-2
This Springer imprint is published by the registered company Springer Nature Switzerland AGThe registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
It is fairly obvious that our world is uncertain, complex, and ambiguous. It is truetoday more than ever, and tomorrow will be more challenging. Artificial intelli-gence (AI) can improve responses to major challenges to our communities fromeconomic and industrial development to health care and disease containment.Nevertheless, artificial intelligence is still growing and improving technology, andthere is a necessity for more creative studies and researches led by both academicsand practitioners.
Artificial Intelligence and Industrial Applications—A2IA’2020, which is the firstedition of an annual international conference organized by the ENSAM—Meknesat Moulay Ismail University, intends to contribute to this common great goal. Itaims to offer a platform for experts, researchers, academics, and industrial practi-tioners working in artificial intelligence and its different applications to discussproblems and solutions, concepts, theories and map out the directions for futureresearch. The connections between institutions and individuals working in this fieldhave to keep growing on and on, and this must have a positive impact on pro-ductivity and effectiveness of researches.
The main topics of the conference were:
– Smart Operation Management– Artificial Intelligence: Algorithms and Techniques– Artificial Intelligence for Information and System Security in Industry– Artificial Intelligence for Energy– Artificial Intelligence for Agriculture– Artificial Intelligence for Health care– Other Applications of Artificial Intelligence
In A2IA’2020 conference proceedings, about 141 papers were received fromaround the world. A total of 58 papers are selected for presentation and publication.In order to maintain a high level of quality, a blind peer review process wasperformed by a large international panel of qualified experts in the conference topicareas. Each submission received at least two reviews, and several received up to
v
five. The papers are evaluated on their relevance to A2IA’2020 tracks and topics,scientific correctness, and clarity of presentation.
The papers are organized in two parts:
– Artificial Intelligence and Industrial Applications:Smart Operation Management (Volume 1)In: Advances in Intelligent Systems and Computing
– Artificial Intelligence and Industrial Applications:Artificial Intelligence Techniques for Cyber-Physical, Digital Twin Systemsand Engineering Applications (Volume 2)In: Lecture Notes in Networks and Systems
We hope that our readers will discover valuable new ideas and insights.Lastly, we would like to express our thanks to all contributors in this book
including those whose papers were not included. We would also like to extend ourthanks to all members of the Program Committee and reviewers, who helped uswith their expertise and valuable time. We are tremendously grateful for the pro-fessional and organizational support from Moulay Ismail University. Finally, ourheartfelt thanks go especially to Springer Nature.
Tawfik MasrourIbtissam El Hassani
Anass Cherrafi
vi Preface
Organization
General Chair
Tawfik Masrour Department of Mathematics and ComputerScience, Artificial Intelligence for EngineeringSciences Team (IASI), Laboratory ofMathematical Modeling, Simulation andSmart Systems (L2M3S), ENSAM, My IsmailUniversity, 50500 Meknes, [email protected]@umi.ac.ma
Co-chairVincenzo Piuri Department of Computer Science University
of Milan via Celoria 18, 20133 Milano (MI),Italy
Keynotes SpeakersAmal El Fallah Seghrouchni Sorbonne University, Paris, FranceMawa Chafii ENSEA, CY Paris University, Paris, FranceAbdellatif Benabdellah University of Le Havre, FranceAli Siadat Arts et Métiers Paris Tech Metz, FranceJiju Antony Heriot-Watt University, Edinburgh, ScotlandAndrew Kusiak University of Iowa, USA
A New Method to Analysis of Internet of Things Malware UsingImage Texture Component and Machine Learning Techniques . . . . . . . 119Saloua Senhaji, Sanaa Faquir, Fidae Harchli, Hajji Tarik,and Mohammed Ouazzani Jamil
Optimization of the SLM Process by Printing a Test Trayin AlSi7Mg06 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126Faraj Zainab, Aboussaleh Mohamed, and Zaki Smail
Dr. Tawfik Masrour is a professor of AppliedMathematics and Artificial Intelligence in NationalHigh School for the Arts and Crafts(ENSAM-Meknes), My Ismaïl University UMI, andmember of Research Team Artificial Intelligence forEngineering Sciences (AIES) and Laboratory ofMathematical Modeling, Simulation and SmartSystems. (L2M3S). He graduated fromMohammed V University – Rabat with MSc degreein Applied Mathematics and, Numerical Analysis, andfrom Jacques-Louis Lions Laboratory, Pierre andMarie Curie University, Paris with M.A.S (DEA) inApplied Mathematics, Numerical Analysis andComputer Science. He obtained his PhD, inMathematics and informatics, from École des PontsParisTech (ENPC), Paris, France. His research inter-ests include Control Theory and Artificial Intelligence.Email : [email protected]
xvii
Dr. Ibtissam El Hassani is a professor of Industrialand Logistic Engineering in National High School forthe Arts and Crafts (ENSAM-Meknès), My IsmaïlUniversity UMI, and head of Research Team ArtificialIntelligence for Engineering Sciences (AIES). Shegraduated as an industrial engineer and then got herPh.D. in Computer Engineering & SystemsSupervision at the Center for Doctoral Studies at MyIsmail University. Her fields of interest are applicationsof artificial intelligence in industrial engineering spe-cially in Lean Manufacturing, Six Sigma &Continuous Improvement, Production systems,Quality, Management System, Health and security atwork, Supply Chain Management, Industrial RiskManagement, and Work Measure.Email: [email protected]
Dr. Anass Cherrafi Is an Assistant Professor at theDepartment of Industrial Engineering, ENSAM-Meknes, Moulay Ismail University, Morocco. Heholds a PhD in Industrial Engineering. He has about7 years of industry and teaching experience. He haspublished a number of articles in leading internationaljournals and conferences. He has participated as GuestEditor for special issues in various internationaljournals. His research interests include industry 4.0,green manufacturing, Lean Six Sigma, integratedmanagement systems and supply chain management.
xviii About the Editors
Efficient Pore Network ExtractionMethod Based on the Distance Transform
Adam Hammoumi1(B), Maxime Moreaud1, Elsa Jolimaitre1,Thibaud Chevalier2, Alexey Novikov3, and Michaela Klotz3
1 IFP Energies nouvelles, Rond-point de l’echangeur de Solaize,BP 3, 69360 Solaize, [email protected]
2 IFP Energies nouvelles, 1 et 4 avenue de Bois-Preau,92852 Rueil-Malmaison, France
3 LSFC Laboratoire de synthese et fonctionnalisation des ceramiques UMR 3080CNRS/SAINT-GOBAIN CREE, Saint-Gobain Research Provence, 550 avenue
Alphonse Jauffret, Cavaillon, France
Abstract. Digital twins of materials allow to achieve accurate predic-tions that help creating novel and tailor-made materials with higher stan-dards. In this paper, we are interested in the characterization of porousmedia. Our attention is drawn to develop a method to describe accu-rately the pore network microstructure of porous materials as presentedin [7]. This work proposes an efficient algorithm based on the distancetransform method [12] which is a widely used method in image pro-cessing. The followed approach suggests that a distance transform map,obtained from a microstructure image, passes through different steps.Starting from local maxima extraction and filtering operation, to end upwith another distance transform with source propagation. We illustrateour algorithm with the well-known Pore Network Model of the litera-ture [13], which supposes that the pore structure is either a networkof connected cylinders or cylinders and spheres. Our approach is alsoapplied on multi-scale Boolean random models modelling complex porousmedia microstructures [11]. The porous media morphological character-istics extracted could be used to simulate complex phenomena as thephysisorption isotherms or other experimental techniques.
Keywords: Distance transform · Local maxima · Porous material ·Multi-scale Boolean model · Pore network model · Physisorption ·Experimental techniques
properties, such as permeability, capillary pressure; and geometrical-topologicalproperties, in particular surface area and interconnectivity of pores. Regardlessany physical consideration, the extraction of internal geometry and topology ofconsidered microstructures is a fundamental step before any numerical simula-tion. However, this process is still ambiguous. We cite from [1] obstacles to thisapproach, such as the lack of concise definitions of what constitutes pore andpore throats in real material with complex geometries. That may have dramaticeffect on the network topology. According to the same reference, the uniquenessof numerical simulation results is not guaranteed for different network models ofthe same material. In fact, in [4], it was demonstrated that different lattice-basedtopologies of the same data sample produced different water retention results.A rough validation of phenomena as the mechanisms of adsorption/desorptioncan be obtained from an over-simplified geometries, such as the ink bottle poreillustrated in [8]. Another widely known method is the pore network model [9]which is used to simulate physical and chemical processes, in particular phaseexchange and non-Darcy flow. The pore network in this type of representations isa discrete grid where the nodes represent the pores and the connections betweenthem constitute the pore necks. These elements are usually represented by sim-ple geometrical elements like spheres or cylinders. The challenge for this type ofapproach is not the representation of the porous network but rather its adap-tation to fit realistic microstructures in terms of simulation results. Anotheralternative is multi-scale stochastic morphological models detailed in [10]. Thesemodels can be used to predict effective material properties of complex geometries[22]. The modelization of the microstructure relies on Boolean random modelsbased on stochastic point processes [2,3,11]. A typical microstructure obtainedby this method contains elements like spheres or platelets [22]. Increasingly com-plex microstructures can be produced by combining different Boolean models;a multi-scale microstructure is generated by performing basic set operations -union and intersection - to several one scale microstructures. In this configura-tion, the solid space is represented by the precedent elements and the porosityspace is deduced from the remaining space.
A fundamental advantage of digital pore network generation methods is com-putation efficiency. A valid formulation of a typical pore network extraction prob-lem is the one that asks: how to discretize the void space into a series of pores andpore necks? In this context, we should differentiate between computer generatedmicrostructures and real microstructures obtained by acquisition, namely X-raytomography. We also keep in mind that the nature of the materials under studyis closely linked to the modeling method. For example, the network extractionalgorithm developed in [5] is unsuitable for highly porous materials. The aimis not to cite all the techniques of pore extraction. Instead we mention someof the closest procedures to our proposal. The first example is skeletonization-based methods, in which the estimation of the medial axe leads to the creationof a graph of pore space. In this approach, pores are placed at the nodes ofthe skeleton and pore–necks are defined by the curved elements of the skeletonthat connect these nodes [6]. Another example is the marker-based watershed
Efficient Pore Network Extraction Method Based on the Distance Transform 3
segmentation that proceeds, after applying a distance transform to the voidspace, to remove all local maxima that lead to an oversegmented watershed [7].In this work, a distance transform based algorithm is proposed. We extract fromthe distance map local maxima. A filtering operation is then applied to avoidoverlapping issues. Then, we cover the void space by the geodesic distance trans-form that yields then again a discretized space of pores. Our approach takes asan input a binary image that represents a porous material microstructure. Thebinary aspect of the image refers to the co-existence of both solid and porosityspace. Our algorithm is called the Pore Network Partitioning method (PNP) andit aims at treating a wide spectrum of microstructures.
2 Method
In this part, for the sake of simplicity, equations and illustrations are mainlyprovided for the 2D case. Associated results can be found in Sect. 3.
2.1 Distance Transform
Consider a 2D Image I represented in the discrete plane E = Z×Z. Each pointP (pixel) is identified by the pair (xp, yp). I is divided into two subsets. Thefirst one is the reference set of points which contains foreground elements. Thesecond one is its complementary. A distance map is an image that evaluates eachelement of I by its distance from the closest point belonging to the reference set.The distance transform is the operation that computes the distance map. It isthe basis of many image processing methods [19]. This transformation relies on aspecifc metric d. We distinguish between exact methods based on the euclideandistance [14,16] and approximate methods based on the chamfer distances [15].The euclidean distance for two points P and Q of the discrete space Z
2 is givenby:
de(P,Q) =√
(xP − xQ)2 + (yP − yQ)2 (1)
In the following, a point will be denoted only by one letter in 2D or 3D. Theeuclidean distance is the exact distance. In Z
3, the equation is given similarlyby:
de(P,Q) =√
(xP − xQ)2 + (yP − yQ)2 + (zP − zQ)2 (2)
However, it can be costly in terms of time complexity. The chamfer distanceswere introduced to accelerate the computation [15]. They evaluate the distancebetween two points as the shortest path according to a pixel-connectivity rela-tion. The chamfer distance between two points x and y can be expressed as:
dc(x, y) = minr∑i
(Widti(xi, xi−1)) (3)
where dti is the distance between two neighboring points xi and xi−1. r is theindex of the target point y. The two approaches differ in terms of the propagation
4 A. Hammoumi et al.
technique considering the weights Wi attributed to local displacements for thechamfer distances. That is, the distance transform is a local transformation. Theoperator of the distance transform reads:
DT d(x) = miny �=P
d(x, y) x, y ∈ E (4)
where the related n-dimensional metric space is denoted (E, d), X is the solidspace and P = E \ X is the porosity space. Given a square orthogonal grid inE, the used algorithm throughout this paper is the raster scanning algorithmwhich is well established in the literature [16]. It consists of a forward and abackward pass. The two passes algorithm ensures obtaining the correct distanceto the nearest foreground elements. This technique is widely used for the problemof connected component labeling [17]. Algorithms based on distance transformcan produce Voronoi Diagrams and Delaunay triangulation [18]. In Fig. (1-a), weshow side by side the original image and its corresponding distance transform. Weobtain a digital representation of the porosity space, where the intensity of thewhite color of background elements is inversely proportional to their proximityto foreground elements. The distance transform can also yield a skeletonizationtransform – an operation in image processing that simplifies an object whileretaining its topology – [20]. The accuracy of the distance transform depends onwhether we use exact or approximate transformations.
2.2 Local Maxima Extraction
To reconstruct the porosity network, we are interested in the extraction of thefarthest points from the foreground elements. These points are extracted andstored. We call this operation Local Maxima Extraction. Each element (pixel orvoxel) value will be compared with other elements of its vicinity. The elementsof maximum value are the ones of interest. The output is therefore sensitive tothe number of the neighbors. The corresponding operator writes as follows:
θ(DT d(x)) =⋃xk
DT d(xk) (5)
for point xk that verify:DT d(xk) ≥ DT d(x′
k) (6)
Where point x′k belongs to the neighborhood V(xk).
In Fig. (1-b), the dots correspond to the local extracted maxima.
2.3 Maxima Filtering
For each maximum point, we create a disk (or a sphere in 3D case) parame-terized by Srk
xk⊂ P , where the radius is the corresponding distance transform
rk = DT d(xk). This step will cover the areas around the extracted maxima.Included elements will have all the same distance transform value. Following the
Efficient Pore Network Extraction Method Based on the Distance Transform 5
(a) (b)
(c)
(d)
Fig. 1. Illustrations of intermediate results of the PNP Method applied to a 2D imageof a Boolean model of spheres [22]. (a) Distance transform. (b) Maxima extractionfrom the obtained distance map. (c) Maxima expansion before filtering. (d) Maximaexpansion after filtering. Each figure contains the initial image on the left and itstransform on the right. Solid space is identified by white color in the original imageand by black color in the transforms. The pore space is initially black and is identifiedby gray-scale colors when it’s discretized. In the (d) transform, a part of pore networkand solid space are both black.
6 A. Hammoumi et al.
creation of disks around maxima points, intermediate disks can be trapped in-between other disks. In this case, we check if there’s an overlapping between theintermediate disk center and its vicinity. Then, at the same position, we comparethe distance transform value and the value of the other disks. If the center valueis lower, then the disk is useless and the center corresponding to the maximumis removed. We write:
∀xk Srkxk
≥ DT d(xk) =⇒ xk = 0 (7)
A subsequent operation of creating disks around the remaining points is thenapplied. In Fig. (2), the center of the black disk is included into a bigger disk.Applying the filter will remove the associated maximum point and its diskaccordingly. Considering intersection of disks, two filtering tecnhiques are intro-duced: standard filtering, shown in Fig. (2-a) and filtering with intersectionsremoval, shown in Fig. (2-b). We define the partition function M and F for eachfiltering technique:
M(x) =
⎧⎪⎨⎪⎩
rk x ∈ (∃ !Srkxk
)rk′ x ∈ A| A =
⋂iS
rixi
, rk′ ≥ ri
0 otherwise
(8)
F (x) =
{rk x ∈ (∃ !Srk
xk)
0 otherwise(9)
M and F refer to standard filtering and filtering with intersections removalrespectively. For M , we keep the intersections between pores. Whereas, for theother technique, we create intermediate pores by removing the intersections.The interest of these filtering methods will be explored further in what follows.Figures (1-c) and (d) illustrate the obtained results before and after the standardfiltering operation. For the 3D case, all the operations described above remainthe same, replacing disks with spheres.
Before After
Standard filtering
(a)
Before After
Filtering with intersections removal
(b)
Fig. 2. Illustrations of the local maxima filtering techniques for the 2D case. (a) Stan-dard filtering. (b) Filtering with intersections removal.
Efficient Pore Network Extraction Method Based on the Distance Transform 7
2.4 Distance Transform with Source Propagation
Previous steps do not yet allow to obtain a partition, some points of the porositybeing unlabeled. A geodesic distance, as illustrated in Fig. (3), measures thelength of the path between two points, this path being constrained to be includedin a given set [21]. The geodesic distance between a point x of P and a subsetX of P is denoted:
DG(x,X) = infx′∈X
DG(x, x′) (10)
P
Q
Fig. 3. The geodesic distance between P and Q is given by the red path, whereas theEuclidean distance is the length of the blue path.
Creation of disks aims not to cover all the porosity space. Filtering operationwith intersections removal creates void as well in-between disks. Therefore, weuse the geodesic distance transform to fill empty spaces. In parallel with thedistance calculation which is carried out step by step, it is possible to propagatethe initial source point. We write:
Sp(x,X) = x′| infx′∈X
DG(x, x′) (11)
The actual propagation of the geodesic distance transforms starts from thepreviously created disks and does not end until all the space is completely filledin. In that manner, we insure covering all the void space. Figures (4-c) and (d)illustrate the pore network partition on a 2D Boolean model of spheres obtainedby the two former filtering techniques. The shapes of pores can be controlled byaltering the used technique. Standard filtering (keeping the intersections) aimsto obtain rounded pore shapes shown in Figs. (4-a) and (c). And filtering withintersections removal produces voronoi-like pore partitions shown in Figs. (4-b)and (d).
8 A. Hammoumi et al.
Fig. 4. Illustrations of the two types of local maxima filtering applied to a 2D image ofa Boolean model spheres. (a) standard filtering and (c) its corresponding pore network.(b) filtering with intersections removal and (d) its corresponding pore network.
3 Discussion
The method is evaluated in terms of: accessibility of pores, realistic modeling ofthe porous network and time complexity. Our procedure allows the extraction ofthe porosity network in few simple steps. The algorithm has no preliminary infor-mation about the spatial organization of the pores inside the porous space, butwill gather information through a series of image transformations. This methodyield a labeled image, where each pore is identified by its size. The accessibil-ity of pores –which is mandatory for accurate simulations– is guaranteed. Thepropagation of the geodesic distance transform permits probing pores with arbi-trary shapes. Realistic pore networks are then modeled. Since we use only twodistance transforms throughout the procedure, the time complexity of both usedalgorithms is linear time O(N) [12]. Therefore, the method generates good inputsfor numerical simulations using simple image transformations at the least com-putation cost. The following pseudo code summarizes the main steps explainedbefore (cf. Algorithm 1).
Our algorithm allows pore network extraction on images of 2003 voxels in1.6 s and bigger images of 5003 in 29 s. The computations were performed usinga personal computer (CPU: intel core i7 2.6 GHz, RAM: 16 GB). For the sake of
Efficient Pore Network Extraction Method Based on the Distance Transform 9
Algorithm 1. PNP method applied to a 2D microstructure image1: procedure Pore Network Partitioning2: I ← Binary Image3: Definition of W, H as the width and the height of I4: T ← W × H5: for n ∈ [0, T ] do6: Computation of :7: Distance map DT d(n); � Distance Transform8: Maxima map Md(n); � Maxima Extraction9: end for
10: for n ∈ [0, T ] do � Preliminary Maxima Expansion11: if Md(n) �= 0 then12: r ← DT d(n)13: Create Disk S(n, r);14: end if15: end for16: for n ∈ [0, T ] do � Maxima Standard Filtering17: if Md(n) �= 0 then18: r ← DT d(n)19: if S(n, r) ≥ DT d(n) then20: Md(j, i) = 021: end if22: end if23: end for24: Repeat Steps 10 �→ 15 � Final Maxima Expansion25: for n ∈ [0, T ] do26: Computation of :27: Geodesic Distance map DG(n); � Pore Network reconstruction28: end for29: end procedure
comparison, the marker-based watershed segmentation algorithm [7] took 200 sfor similar sized images of 5003 voxels with the same kind of personal computer.
We provide our results for two types of micro-structures. The first one isa 2D pore network model taken from [13]. It is made of interconnected cylin-ders assembling pores and pores necks. The parameterization of this networkis relatively simple. Each cylinder is identified by its length and diameter. Theintersections of cylinder’s endpoints constitute the pore network. In Fig. (5-b),we show an example of 8002 pixels image.
A multi-scale microstructure is also being considered. It is based on the multi-scale cox Boolean models [11]. To generate this type of microstructure, we fix thesimulation parameters for the aggregates and for the grains; size of aggregates,size of grains and grain volume fraction Vv. The multi-scale aspect is beingdefined by: the aggregates, the grains inside and the outer space. We study thecase of a multi-scale Boolean model of platelets represented in a volume of 2003
voxels. A platelet is defined by its length, height and thickness (L = 25,H =15, T = 5), the rest of the parameters are: Vvin = 0.2, Vvout = 0.4, Rinc = 100
10 A. Hammoumi et al.
(a) (b)
Fig. 5. Illustration of the PNP method applied to a 2D image of pore network model.(a) The original microstructure (white: solid space, black: porosity space). (b) Corre-sponding partition pore network (grayscale elements: pores, black elements: solid).
(radius of the sphere modeling the aggregates), Vvinc = 0.4. The PNP relativeresult is shown in Fig. (6). Partition labeling (originated from a mapping fromthe distance transform to a greyscale image) is provided.
(a)
0
Partition label
3
5
10
12
35
36
39
(b)
Fig. 6. Illustrations of the extracted pore networks by the PNP method. (a) 3D Porenetwork of a multi-scale Boolean model of platelets. (b) Section of the pore network(a) with partition labeling.
A porous material is characterized by descriptors as the volume fraction ofpores, surface area, pore size distribution, etc. These parameters are obtainedexperimentally but could also be simulated numerically. This method should be
Efficient Pore Network Extraction Method Based on the Distance Transform 11
0 5 10 15 20 25 30 35 40
1
2
3
4
5
6·106
Pore size [voxels]
Cum
ulativesu
m[N
umber
ofvo
xels]
PSD
Gr
Fig. 7. Pore Size Distribution of the microstructure shown in Fig. (6-a) (complementaryof multi-scale Boolean model of platelets).
regarded as a first brick for numerical simulations of porous material properties.For instance, it can be used to compute the pore size distribution.
Consider the family of sets (each set is made of pores), and every set has aunique label (pore diameter � DT d). We consider the family of sets {Lr} suchthat r = 1, 2, ....n and n = max(DT d). A Granulometry function is obtainedwith:
Gr = Cardinal(Lr) (12)
The pore size distribution ‘PSD’ is the cumulative sum of the granulometryfunction values and is defined by:
H0 = 0Hr+1 = Gr+1 + Hr where r = 0, 1, ...n − 1 (13)
An example of the PSD computation is shown in Fig. (7).The interested reader can find the PNP algorithm on the open access software
“plug im!” website, https://www.plugim.frThe Figs. (1), (4), (5), (6) were generated using “plug im!” (2020).
4 Conclusion
A new algorithm for pore network extraction has been described. The algorithmrelies on simple and well-known methods in image processing. The method is
a mapping from the continuum void space contained initially in a computer-generated microstructure to a discretized space made of distinct pores in termsof size and shape. In particular, the method is suitable for complex microstruc-tures as shown above, but also for discrete grids made of interconnected spheresand/or cylinders. At this stage, the method is potentially interesting due to itssimplicity and computational efficiency. The labeled extracted pore network fitsnumerical simulations requirements. It makes it possible to manage pores andtheir connectivities. Future work will focus on enhancing the accuracy of theproposed algorithm. Namely, by the use of efficient neural network algorithms.The simulation of physico-chemical processes based on the data produced by thePNP algorithm is the next step to validate the method.
References
1. Bhattad, P., Willson, C.S., Thompson, K.E.: Effect of network structure on char-acterization and flow modeling using x-ray micro-tomography images of granularand fibrous porous media. Transp. Porous Med. 90, 363 (2011)
3. Matheron, G.: Random Sets and Integral Geometry. Wiley, New York (1975)4. Vogel, H.J.: A numerical experiment on pore size, pore connectivity, water reten-
tion, permeability, and solute transport using network models. Eur. J. Soil Sci.51(1), 99–105 (2000)
5. Yang, F., Hingerl, F., Xiao, X., et al.: Extraction of pore-morphology and capillarypressure curves of porous media from synchrotron-based tomography data. Sci.Rep. 5, 10635 (2015)
6. Pudney, C.: Distance-based skeletonization of 3D images. In: Proceedings of the1996 IEEE TENCON Digital Signal Processing Applications, TENCON 1996, vol.201, pp. 209–214 (1996)
7. Gostick, J.T.: Versatile and efficient pore network extraction method using marker-based watershed segmentation. Phys. Rev. E 96, 023307 (2017)
8. Zeng, Y., Fan, C., Do, D.D., Nicholson, D.: Evaporation from an ink-bottle pore:mechanisms of adsorption and desorption. Ind. Eng. Chem. Res. 53(40), 15467–15474 (2014)
9. Xiong, Q., Baychev, T.G., Jivkov, A.P.: Review of pore network modelling ofporous media: experimental characterisations, network constructions and appli-cations to reactive transport. J. Contam. Hydrol. 192, 101–117 (2016)
10. Jeulin, D.: Morphology and effective properties of multi-scale random sets: areview. Comptes Rendus Mecanique 340(4–5), 219–229 (2012)
11. Moreaud, M., Chaniot, J., Fournel, T., Becker, J.M., Sorbier, L.: Multi-scalestochastic morphological models for 3D complex microstructures. In: IEEE Con-ference on 17th Workshop on Information Optics (WIO), Quebec city, Canada(2018)
12. Maurer Jr., C.R., Qi, R., Raghavan, V.: A linear time algorithm for computingexact Euclidean distance transforms of binary images in arbitrary dimensions.IEEE Trans. Pattern Anal. Mach. Intell. 25(2), 265–270 (2003)
13. Ledezma, G., Verstraete, J., Sorbier, L., Leinekugel-Le Cocq, D., Jolimatre, E.,Jallut, C.: Computational characterization techniques applied to a pore networkmodel by using a fast percolation algorithm (article under preparation, 2020)
Efficient Pore Network Extraction Method Based on the Distance Transform 13
14. Bailey, D.G.: An efficient Euclidean distance transform. In: Klette R., Zunic J.(eds.) Combinatorial Image Analysis, IWCIA 2004. LNCS, vol. 3322, pp. 394–408.Springer, Berlin (2004)