Sébastien Drouyer: PhD at the CMM, MINES ParisTech

My name is Sébastien Drouyer and I am currently finishing a PhD thesis in the Center for Mathematical Morphology (CMM) at MINES Paristech, under the supervision of Serge Beucher and Michel Bilodeau. The PhD is financed by IFP Energies Nouvelles where I also work with Maxime Moreaud and Loïc Sorbier.

Mathematical morphology is a field of image processing. Since its creation, it has developed various mathematical tools to process shapes and geometrical structures in the images (hence the term “morphology”).

Mathematical Morphology has been created by Georges Matheron and Jean Serra at the Ecole des Mines de Paris, sharing a strong background on both mathematics and geology. This specificity brings an original approach to image processing: the image is viewed as a topographic map and most operators have been developed with this perspective in mind. The origin of Mathematical Morphology and its novel approach has allowed its researchers to apply its concepts to multiple fields of study: image processing of course, but also lattice theory, stochastic geometry, nonlinear PDE, graph theory, etc. as well as the application to the study of materials and physical simulations.

The objective of my PhD thesis is to create an automatic process that estimates the 3D topography of a microscopic sample from multiple images taken using a Scanning Electron Microscope. The process of obtaining depth information from 2D images is called stereo reconstruction, and an extensive literature exists on the subject. There is a wide range of applications: cartography, robotics and autonomous driving constitute only a small fraction of existing applications.

Estimating the 3D topography of microscopic samples from SEM images

However, the state of the art methods do not work well on our microscopic samples due to the presence of low textured areas. We have therefore developed a novel approach combining existing stereo methods and hierarchical segmentation algorithms that divide the images in regions and sub-regions. This general approach allows us to apply our method to any stereo images, exceeding the initial specific application to SEM images of my thesis.

Results obtained so far are promising: we obtained results that are comparable and sometimes better than state of the art methods in the Middlebury online ranking (our submission’s name is MC-CNN+TDSR) that ranks more than 60 stereo methods on a standard stereo images dataset. On our SEM images, results obtained are also of better quality (see Figure).

I have additionally won prizes in the IARPA stereo challenge, where contestants had to create a software that automatically estimates the 3D mapping of urban areas from satellite images.

Here is my personal webpage: http://sebastien.drouyer.com.

%d bloggers like this: