Keynote Speakers

Stéphane Roux
Director of Research, CNRS

Stéphane Roux graduated from the Ecole Polytechnique in 1983 and the Ecole Nationale des Ponts et Chaussées (ENPC) in 1985. He received his Ph.D. in mechanical engineering from the ENPC in 1990. As a CNRS Research Professor, he served successively at the Ecole Supérieure de Physique et Chimie Industrielles de la Ville de Paris (ESPCI), at the joint CNRS/Saint-Gobain Research Laboratory, and currently, he is at the Laboratory of Mechanics Paris-Saclay at the Ecole Normale Supérieure de Paris-Saclay. His research activity is devoted to data processing and image-based measurements for experimental mechanics. This includes digital image correlation, stereo-correlation (for surface reconstructions in 3D), and also digital volume correlation for tomography. He holds 17 patents and is the author of more than 420 publications, (H=86, Google Scholar). He received the Silver Medal from the CNRS in 2006, and the Jaffé prize (French Academy of Sciences) in 2019.

Abstract: Yarn-path extraction in large 3D textiles based on periodicity and volume registration

Segmenting yarn paths in 3D woven reinforcements from low-resolution tomographic images is notoriously difficult and time-consuming, and is still performed manually today. The chosen example is the segmentation of warp and weft yarns at the root of a Leap fan blade, with a voxel size of 140 µm, and the image extending over the entire part.

The initial observation is that, in some regions, the weaving topology is periodic. From an autocorrelation analysis, a unit periodic cell is easily identified, and, by exploiting its periodicity, warp and weft yarns are segmented within the cell. This task is accomplished using local matching with a template of the yarn cross-section, exploiting the continuity of each yarn path and the consistency of the warp and weft paths in 3D space, with no overlap. The unit cell is tiled along the three space directions to create a reference volume of arbitrary size. Global Digital Volume Correlation registers the actual image with the periodic reference, accounting for large-scale distortions to match the two 3D images. The global approach is based on a fine-mesh discretization of the displacement field and mechanical regularization. These two features help avoiding local-minima trapping, as can be feared in an almost periodic medium. The mapping can now be used to transfer the unit-cell segmentation to the entire volume, thereby providing a sound labeling of each individual yarn.

In contrast to many other alternatives, this procedure does not rely on numerous manual annotations, does not require any training, and additionally provides reliability indicators based on registration residuals that highlight any misregistration.