PhD Position in Machine Learning and Stochastic Geometrical Image Analysis
ON THE COMBINATION OF MACHINE LEARNING AND STOCHASTIC GEOMETRY FOR IMAGE ANALYSIS OF PARTICLE POPULATIONS
Supervisor: Prof. Johan Debayle (MINES Saint-Étienne - SPIN / UMR CNRS 5307 LGF, France) Collaborations: University of Leon (Spain), University of Burgundy (France), CEA Marcoule (France) Location: MINES Saint-Etienne (France)
Starting date / Duration: october 2019 / 36 months Context
Most of the steps of process/chemical engineering involve multiphase flows, like liquid-liquid extraction, crystallization, filtration, precipitation, etc. For all of these applications the knowledge of the main properties of the dispersed phase (particle shape, particle size distribution, etc.) is a key issue. In this way, image acquisition and analysis can be used to estimate the characteristics of these particles (drops, bubbles, crystals…). Nevertheless, these quantitative measurements are generally based on restrictive assumptions (low density of particles, isotropic particles, etc.) which represent a strong limitation for the scope of investigation.
Research works for analyzing populations of particles with a high density have shown the limitations of low- level image processing methods based on particles individualization and characterization. To overcome these limitations, approaches based on stochastic geometry have been investigated. Nevertheless, for non-spherical shapes, the optimization process to fit the stochastic geometrical model to the real data is not straightforward (cost function, anisotropy, a priori shape of the particles…). In order to handle this problem, machine learning approaches are a promising alternative.
The main objective of this PhD thesis is the development of methods combining machine learning and stochastic geometry for characterizing the geometry and morphology of a dense population of particles, adapted to the encountered configurations (bubbles, droplets, crystals ...).
(a) crystals (b) liquid-gaz flow (c) emulsion
State of the art
Regarding the literature, low-level image processing methods for analyzing such 2-D images are based on mathematical morphology [1], Hough transform [2] and pattern recognition [3-8]. Such methods revealed good performances for the detection of particles in several processes. However, they reach their limits and are not appropriate to the most demanding configurations encountered in R&D studies (viscous liquids, bubbles, anisotropic solids), where the inclusion of higher density particles (involving many overlapping with the 2-D projections) as well as specific morphologies (e.g. elongated objects) require specific developments. More specifically, these methods are mainly based on particle individualization, so it becomes a difficult task when the overlapping of projected particles is high. To overcome these limitations, stochastic geometrical methods [9-11] have been investigated for such applications [12]. The purpose is to geometrically model the population of particles with random sets so as to simulate synthetic images representative of the real data. They address some limitations related to the population density by providing, without any step of object individualization, the geometrical characteristics of individual particles (average area, average perimeter, number) [13, 14].
However, outside the well-known Boolean model (which is generally not representative of physical situations), there is much less theoretical results. For more flexible stochastic geometrical models (Matern, Quermass, Gibbs…), a numerical optimization process is required to fit the model. The main difficulty of this step concerns the cost function: its choice is not straightforward and can provide different solutions to the optimization problem. Alternatively, machine learning approaches (support vector machines, artificial neural networks, Bayesian networks, convolutional neural networks) [15] can be used to make this optimization process.
Objective and working plan
The aim of this work is to develop and study stochastic geometrical methods for characterizing the geometry and the morphology of dense particle populations such as concentrated emulsions, flows of non-spherical bubbles and solid particles encountered in chemical engineering processes.
After a literature review, stochastic geometry tools will be proposed and implemented to characterize images of dense particle populations from actual cases. For this purpose synthetic images will be simulated using a stochastic model for which both its geometrical and morphological parameters (density, size distribution, shape, orientation ...) should be adjusted by statistical comparison to the real images via machine learning tools. For each configuration (bubbles, crystals, etc.), the methodology will be based on the following steps:
Development of a stochastic model (Boolean model, dead leaves, random field, etc.) and simulation of synthetic images,
Characterization of synthetic images, i.e. extraction of descriptors (total area, covariance, Euler number, etc.) for different parameter configurations,
Identification of the model parameters by using machine learning,
Application to real images (a database of crystal images could be firstly provided).
The result of the parameters identification will provide geometrical (area), topological (number of objects) and morphological (shape factor) characteristics of the individual particles in a statistical sense.
Finally, an important part of the study will consist in proposing, for each configuration, a validation system based on numerical and/or experimental studies, so as to evaluate the performance of the proposed methods.
For this purpose, the important database already available can be used.
Bibliography
[1] A. Khalil, F. Puel, Y. Chevalier, J. M. Galven, A. Rivoire, J. P. Klein, Chem. Eng. J. 2012, 165, 946.
[2] Hans-Jörg Bart, Matthias Mickler, and Hanin B. Jildeh. Optical Image Analysis and Determination of Dispersed Multi Phase Flow for Simulation and Control in Optical Imaging: Technology, Methods and Applications. Nova Science Publishers, 2012.
[3] S. Maaß, J. Rojahn, R. Hänsch and M. Kraume. Automated drop detection using image analysis for online particle size monitoring in multiphase systems. Computers & Chemical Engineering, 45: 27 – 37, 2012.
[4] O. Ahmad, J. Debayle, N. Gherras, B. Presles, G. Fevotte, and J. C. Pinoli. Quantification of overlapping polygonal-shaped particles based on a new segmentation method of in situ images during crystallization.
Journal of Electronic Imaging, 21(2):1-12, 2012.
[5] B. Presles, J. Debayle, and J. C. Pinoli. Size and shape estimation of 3-D convex objects from their 2-D projections. Application to crystallization processes. Journal of Microscopy, 248(2):140-155, 2012.
[6] O. Ahmad, J. Debayle, and J. C. Pinoli. A geometric-based method for recognizing overlapping polygonal- shaped and semi-transparent particles in gray tone images. Pattern Recognition Letters, 32(15):2068-2079, 2011.
[7] B. Presles, J. Debayle, G. Fevotte and J. C. Pinoli. A novel image analysis method for in-situ monitoring the particle size distribution of batch crystallization processes. Journal of Electronic Imaging, 19(3):1-7, 2010.
[8] M. De Langlard, H. Al Saddik, S. Charton, J. Debayle, and F. Lamadie. An efficiency improved recognition algorithm for highly overlapping ellipses: Application to dense bubbly flows. Pattern Recognition Letters, 101:88-95, 2018.
[9] S.N. Chiu, D. Stoyan, W.S. Kendall and J. Mecke. Stochastic geometry and its applications. Wiley, 2013.
[10] I. S. Molchanov. Statistics of the Boolean Model for Practitioners and Mathematicians. Wiley, Chichester 1997.
[11] K. Michielsen and H. De Raedt. Integral geometry morphological image analysis. Physical Reports, 347:461-538, 2001.
[12] M. De Langlard, F. Lamadie, S. Charton, and J. Debayle. A 3D stochastic model for geometrical characterization of particles in two-phase flow applications. Image Analysis and Stereology, 37(3):233-247, 2018.
[13] R. Schneider and W. Weil. Stochastic and Integral Geometry. Springer, Berlin Heidelberg, 2008.
[14] M. Berchtold. Modelling of Random Porous Media using Minkowski Functionals. PhD Thesis, ETH Zurich, Swiss, 2008.
[15] C.M. Bishop. Pattern Recognition and Machine Learning, Springer, 2006.
Candidate profile
Candidates should have a MSc. in applied mathematics and/or computer science: image analysis, stochastic geometry, machine learning
Programming skills with Matlab/Python and C/C++
A good level of written and spoken English Contact
Candidates should send cover letter, CV and reference letters to Prof. Johan DEBAYLE (debayle@emse.fr).
