Search

International audience; Bleomycin (BLM) is a potent anticancer drug used to treat different malignancies, mainly lymphomas, germ cell tumors, and melanomas. Unfortunately, BLM has major, dose-dependent, pulmonary toxicity that affects 20% of treated individuals. The most severe form of BLM-induced pulmonary toxicity is lung fibrosis. Deglyco-BLM is a molecule derived from BLM in which the sugar residue D-mannosyl-L-glucose disaccharide has been deleted. The objective of this study was to assess the anticancer activity and lung toxicity of deglyco-BLM. We compared the antitumor activity and pulmonary toxicity of intraperitoneally administrated deglyco-BLM and BLM in three rodent models. Pulmonary toxicity was examined in depth after intratracheal administration of both chemotherapeutic agents. The effect of both drugs was further studied in epithelial alveolar cells in vitro. We demonstrated in rodent cancer models, including a human Hodgkin's lymphoma xenograft and a syngeneic melanoma model, that intraperitoneal deglyco-BLM is as effective as BLM in inducing tumor regression. Whereas the antitumor effect of BLM was accompanied by a loss of body weight and the development of pulmonary toxicity, deglyco-BLM did not affect body weight and did not engender lung injury. Both molecules induced lung epithelial cell apoptosis after intratracheal administration, but deglyco-BLM lost the ability to induce caspase-1 activation and the production of ROS (reactive oxygen species), transforming growth factor-beta 1, and other profibrotic and inflammatory cytokines in the lungs of mice and in vitro. Deglyco-BLM should be considered for clinical testing as a less toxic alternative to BLM in cancer therapy.

International audience; We propose a new computer aided detection framework for tumours acquired on DCE-MRI (Dynamic Contrast Enhanced Magnetic Resonance Imaging) series on small animals. In this approach we consider DCE-MRI series as multivariate images. A full multivariate segmentation method based on dimensionality reduction, noise filtering, supervised classification and stochastic watershed is explained and tested on several data sets. The two main key-points introduced in this paper are noise reduction preserving contours and spatio temporal segmentation by stochastic watershed. Noise reduction is performed in a special way that selects factorial axes of Factor Correspondence Analysis in order to preserves contours. Then a spatio-temporal approach based on stochastic watershed is used to segment tumours. The results obtained are in accordance with the diagnosis of the medical doctors.

Recent road crack detection methods obtain appealing scores but typically allow a few pixel tolerance margin. This is acceptable for locating cracks, but not for measuring their width (indicator of the cracks' severity). Our baseline model, U-VGG19, obtains an F-score of 71.77% on CrackForest, which is superior to other approaches when no tolerance is admitted. However, increasing the scores without tolerance is difficult due to inaccurate annotations. We propose a novel synthetic dataset, Syncrack, as a benchmark for the evaluation of training with inaccurate annotations. Our results show that inaccurate annotations have a detrimental impact on the F-measure, decreasing it by up to 20%. To overcome this, we study label noise correction techniques using weakly supervised learning. Training U-VGG19 with these corrected labels improves the results on Syncrack by up to 12%. Obtained results on the CrackForest and Aigle-RN datasets support that these approaches are useful for real-life data too.

International audience; The extrinsic calibration refers to determining the relative pose of cameras. Most of the approaches for cameras with non-overlapping fields of view (FOV) are based on mirror reflection, object tracking or rigidity constraint of stereo systems whereas cameras with overlapping FOV can be calibrated using structure from motion solutions. We propose an extrinsic calibration method within structure from motion framework for cameras with overlapping FOV and its extension to cameras with partially non-overlapping FOV. Recently, omnidirectional vision has become a popular topic in computer vision as an omnidirectional camera can cover large FOV in one image. Combining the good resolution of perspective cameras and the wide observation angle of omnidirectional cameras has been an attractive trend in multi-camera system. For this reason, we present an approach which is applicable to heterogeneous types of vision sensors. Moreover, this method utilizes images of lines as these features possess several advantageous characteristics over point features, especially in urban environment. The calibration consists of a linear estimation of orientation and position of cameras and optionally bundle adjustment to refine the extrinsic parameters.

International audience; In continuum mechanics, the prediction of defect harmfulness requires to solve approximately partial differential equations with given boundary conditions. In this contribution boundary conditions are learnt for tight local volumes (TLV) surrounding cracks in 3D volumes. A non-parametric data-driven approach is used to define the space of defects, by considering defects observed via X-Ray computed tomography (Lacourt et al, 2020; Ryckelynck et al, 2020). The dimension of the ambient space for the observed images of defects is huge. A nonlinear dimensionality reduction scheme is proposed in order to train a reduced latent space for both the morphology of defects and their local mechanical effects in the TLV. A multimodal autoencoder (Bhatt et al, 2019) enables to mix morphological and mechanical data. It contains a single latent space, termed mechanical latent space. But this latent space is fed by two encoders. One is related to the images of defects and the other to mechanical fields in the TLV. The latent variables are input variables for a geometrical decoder and for a mechanical decoder. In this work, mechanical variables are displacement fields. The autoencoder on mechanical variables enables projection-based model order reduction as proposed in (Lee et al, 2020). The main novelty of this paper is a submodeling approach assisted by artificial intelligence. Here, for defect images in the test set, Dirichlet boundary conditions are applied to TLV. These boundary conditions are forecasted by the mechanical decoder with a latent vector predicted by the morphological encoder. For that purpose, a mapping is trained to convert morphological latent variables into mechanical latent variables, denoted ``direct mapping''. An ``inverse mapping'' is also trained for error estimation with respect to morphological predictions. Errors on mechanical predictions are close to 5% with simulation speed-up ranging for 3 to 120. We show that latent variables forecasted by the images of defects are prone to a better understanding of the predictions.

International audience; Most recent methods of image augmentation and prediction are building upon the deep learning paradigm. A careful preparation of the image dataset and the choice of a suitable network architecture are crucial steps to assess the desired image features and, thence, achieve accurate predictions. We first propose to help the learning process by adding structural information with specific distance transform to the input image data. To handle cases with limited number of training samples, we propose a patch-based procedure with a stratified sampling method at inference. We validate our approaches on two image datasets, corresponding to two different tasks. The ability of our method to segment and predict images is investigated through the ISBI 2012 segmentation challenge dataset and generated electric field masks, respectively. The obtained results are evaluated using appropriate metrics: VRand for image segmentation and SSIM, UIQ and PSNR for image prediction. The proposed techniques demonstrate that the established framework is a reliable estimation method that could be used for a wide range of applications.

International audience; This paper proposes a novel approach of line matching across images captured by different types of cameras, from perspective to omnidirectional ones. Based on the spherical mapping, this method utilizes spherical SIFT point features to boost line matching and searches line correspondences using an affine invariant measure of similarity. It permits to unify the commonest cameras and to process heterogeneous images with the least distortion of visual information.

International audience; In the last 20 years, 3D angiographic imaging has proven its usefulness in the context of various clinical applications. However, angiographic images are generally difficult to analyse due to their size and the complexity of the data that they represent, as well as the fact that useful information is easily corrupted by noise and artifacts. Therefore, there is an ongoing necessity to provide tools facilitating their visualisation and analysis, while vessel segmentation from such images remains a challenging task. This article presents new vessel segmentation and filtering techniques, relying on recent advances in mathematical morphology. In particular, methodological results related to spatially variant mathematical morphology and connected filtering are stated, and included in an angiographic data processing framework. These filtering and segmentation methods are evaluated on real and synthetic 3D angiographic data.

X-ray Computed Tomography (CT) has been increasingly used in many industrial domains for its unique capability of controlling both the integrity and dimensional conformity of parts. Still, it fails to be adopted as a standard technique for on-line monitoring due to its excessive cost in terms of acquisition time. The reduction of the number of projections, leading to the so-called sparse-view CT strategy, while maintaining a sufficient reconstruction quality is therefore one of the main challenges in this field. This work aims to evaluate and compare the performances of two deep learning strategies for the sparse-view reconstruction problem. As such, we propose an extensive study of these methods, both in terms of data regime and angular sparsity during training. The two strategies present quantitative improvements over a classical FBP/FDK approach with a PSNR improvement varying between 11 and 16 dB (depending on the angular sparsity) ; showing that efficient CT inspection can be performed from only few dozens of images

International audience

ISBN : 978-1-4244-2002-5; International audience; Segmenting dynamic contrast enhanced-MRI series of small animal, which are intrinsically noisy and low contrasted images with low resolution, is the aim of this paper. To do this, a segmentation method taking into account the temporal (spectral) and spatial information is presented on several series. The idea is to start from a temporal classification, and to build a probability density function of contours conditionally to this classification. Then, this function is segmented to find potentially tumorous areas. The method is presented on several series after a range normalization histogram in order to compare the series.

Homologous Recombination DNA-repair deficiency (HRD) is a well-recognized marker of platinum-salt and PARP inhibitor chemotherapies in ovarian and breast cancers (BC). Causing high genomic instability, HRD is currently determined by BRCA1/2 sequencing or by genomic signatures, but its morphological manifestation is not well understood. Deep Learning (DL) is a powerful machine learning technique that has been recently shown to be capable of predicting genomic signatures from stained tissue slides. However, DL is known to be sensitive to dataset biases and lacks interpretability. Here, we present and evaluate a strategy to control for biases in retrospective cohorts. We train a deep-learning model to predict the HRD in a controlled cohort with unprecedented accuracy (AUC: 0.86) and we develop a new visualization technique that allows for automatic extraction of new morphological features related to HRD. We analyze in detail the extracted morphological patterns that open new hypotheses on the phenotypic impact of HRD.

The original publication is available at www.ias-iss.org.; International audience; Morphological characterization of porous media is of paramount interest, mainly due to the connections between their physicochemical properties and their porous microstructure geometry. Heterogeneity can be seen as a geometric characteristic of porous microstructures. In this paper, two novel topological descriptors are proposed, based on the M-tortuosity formalism. Using the concept of geometric tortuosity or morphological tortuosity, a first operator is defined, the H-tortuosity. It estimates the average variations of the morphological tortuosity as a function of the scale, based on Monte Carlo method and assessing the heterogeneity of porous networks. The second descriptor is an extension, named the H-tortuosity-by-iterative-erosions, taking into account different percolating particle sizes. These two topological operators are applied on Cox multi-scale Boolean models, to validate their behaviors and to highlight their discriminative power.The original publication is available at www.ias-iss.org.

International audience; The clear description of the morphology and location, with respect to the support, of metallic sub-nanometric particles remains a current experimental strenuous challenge in numerous catalytic applications. High resolution-HAADF-STEM coupled with in situ and tomographic analyses are undertaken on platinum (Pt) active phase supported on chlorinated alumina (γ-Al 2 O 3) with 0.3 and 1% w/w Pt loadings highlighting the formation of flat nanoparticles (NPs) of 0.9 nm diameter and Pt single atoms (SAs) in the reduced state. While SAs and weakly cohesive clusters are predominantly observed in the oxide state, with a coordination sphere of Pt composed of O and Cl as revealed by EXAFS, the ratio between SAs and Pt NPs in the reduced state is found to be about 2.8. This ratio is not affected by metal loading which increases both the total number of NPs and SA. Electron tomography reveals that the vast majority of NPs are located on the edges or defects (steps, kinks) of the alumina support crystallites. Density functional theory calculations further highlight the optimized structures of NPs located at the γ-Al 2 O 3 (110)(100) edge and near-edge with a stability competing with NPs located either on the (110) or on the (100) γ-Al 2 O 3 facet. A mathematical analysis of the segmented volumes shows that the average geodesic distances between NPs is linked to Pt loading: 9 nm for 1% w/w Pt, and 16 nm for 0.3% w/w Pt. Evaluation of support tortuosity descriptors using the nanoparticles positions confirms a uniform distribution on the support. A square network geometric model compatible with the geodesic distances between NPs reveals that 1 to 5 NPs can be present at the same time on each alumina crystallite depending on Pt loading.

International audience; The methods used to extract chord-length distributions from small-angle scattering data assume a structure consisting of spatially uncorrelated and disconnected convex regions. These restrictive conditions are seldom met for a wide variety of materials such as porous materials and semicrystalline or phase-separated copolymers, the structures of which consist of co-continuous phases that interpenetrate each other in a geometrically complex way. The significant errors that would result from applying existing methods to such systems are discussed using three distinct models for which the chord-length distributions are known analytically. The models are a dilute suspension of hollow spheres, the Poisson mosaic and the Boolean model of spheres.

Simple probabilistic models of fracture of 3D polycrystals involving transgranular and intergranular cracks are proposed, based on undamaged paths and appropriate percolation thresholds. In a second part, we propose theoretical extensions of phase field models for crack initiation and propagation to the case of locally heterogeneous and anisotropic fracture energy, and their implementation to get full field solutions by means of iterations of Fourier transforms to replace the standard finite element approach.

The segmentation of tomographic images of the battery electrode is a crucial processing step, which will have an additional impact on the results of material characterization and electrochemical simulation. However, manually labeling X-ray CT images (XCT) is time-consuming, and these XCT images are generally difficult to segment with histographical methods. We propose a deep learning approach with an asymmetrical depth encode-decoder convolutional neural network (CNN) for real-world battery material datasets. This network achieves high accuracy while requiring small amounts of labeled data and predicts a volume of billions voxel within few minutes. While applying supervised machine learning for segmenting real-world data, the ground truth is often absent. The results of segmentation are usually qualitatively justified by visual judgement. We try to unravel this fuzzy definition of segmentation quality by identifying the uncertainty due to the human bias diluted in the training data. Further CNN trainings using synthetic data show quantitative impact of such uncertainty on the determination of material’s properties. Nano-XCT datasets of various battery materials have been successfully segmented by training this neural network from scratch. We will also show that applying the transfer learning, which consists of reusing a well-trained network, can improve the accuracy of a similar dataset.

International audience; In crack detection, pixel-accurate predictions are necessary to measure the width-an important indicator of the severity of a crack. However, manual annotation of images to train supervised models is a hard and time-consuming task. Because of this, manual annotations tend to be inaccurate, particularly at pixel-accurate level. The learning bias introduced by this inaccuracy hinders pixel-accurate crack detection. In this paper we propose a novel tool aimed for synthetic image generation with accurate crack labels-Syncrack. This parametrizable tool also provides a method to introduce controlled noise to annotations, emulating human inaccuracy. By using this, first we do a robustness study of the impact of training with inaccurate labels. This study quantifies the detrimental effect of inaccurate annotations in the final prediction scores. Afterwards, we propose to use Syncrack to avoid this detrimental effect in a real-life context. For this, we show the advantages of using Syncrack generated images with accurate annotations for crack detection on real road images. Since supervised scores are biased by the inaccuracy of annotations, we propose a set of unsupervised metrics to evaluate the segmentation quality in terms of crack width.

International audience; Morphological models are commonly used to describe microstructures observed in heterogeneous materials. Usually, these models depend upon a set of parameters that must be chosen carefully to match experimental observations conducted on the microstructure. A common approach to perform the parameters determination is to try to minimize an objective function, usually taken to be the discrepancy between measurements computed on the simulations and on the experimental observations, respectively. In this article, we present a Bayesian approach for determining the parameters of morphological models, based upon the definition of a posterior distribution for the parameters. A Monte Carlo Markov Chains (MCMC) algorithm is then used to generate samples from the posterior distribution and to identify a set of optimal parameters. We show on several examples that the Bayesian approach allows us to properly identify the optimal parameters of distinct morphological models and to identify potential correlations between the parameters of the models.

The work deals with a macroscopically isotropic energetic material based on triamino-trinitrobenzene (TATB) crystals bonded with a small volume fraction of a thermoplastic polymer. This material is shown experimentally to display an irreversible thermal expansion behavior characterized by dilatancy and variations of its thermal expansion coefficient when heated or cooled outside a narrow reversibility temperature range. The analysis of cooling results suggests the existence of residual stresses in the initial state, attributed to the manufacturing process. Microstructure-level FFT computations including the very strong anisotropic thermoelastic TATB crystal response and temperature-dependent binder plasticity, show that strong internal stresses develop in the disoriented crystals under thermal load, either heating or cooling. Upon cooling, binder plastic yielding in hindered, thus promoting essentially brittle microcracking, while it is favored upon heating. Despite its low volume fraction, the role of the binder is essential, its plastic yielding causing stress redistribution and residual stresses upon cooling back to ambient.

International audience; La sensibilité à l'amorçage d'un explosif est intimement lié à sa microstructure. La présence d'hétérogénéités au sein de l'explosif (phases, pores, fissures, etc.) provoque, sous chargement mécanique (traction-compression, choc), des déformations localisées et donc des échauffements localisés, ou points chauds, sites préférentiels de démarrage des réactions chimiques. Pour étudier la sensibilité à l'amorçage d'un explosif, il est donc nécessaire de caractériser les hétérogénéités de sa microstructure.

Point clouds have attracted the interest of the research community over the last years. Initially, they were mostly used for remote sensing applications. More recently, thanks to the development of low-cost sensors and the publication of some opensource libraries, they have become very popular and have been applied to a wider range of applications. One of them is the autonomous vehicle where many efforts have been made in the last century to make it real. A very important bottleneck nowadays for the autonomous vehicle is the evaluation of the proposed algorithms. Due to the huge number of possible scenarios, it is not feasible to perform it in real life. An alternative is to simulate virtual environments where all possible configurations can be set up beforehand. However, they are not as realistic as the real world is. In this thesis, we studied the pertinence of including hyperspectral images in the creation of new virtual environments. Furthermore, we proposed new methods to improve 3D scene understanding for autonomous vehicles.; Les nuages de points ont suscité l'intérêt de la communauté de recherche au cours des dernières années. Au départ, ils étaient surtout utilisés pour des applications de télédétection. Plus récemment, grâce au développement de capteurs à faible coût et à la publication de plusieurs bibliothèques open source, ils sont devenus très populaires et ont été appliqués à un plus grand nombre d'applications. L'une d'entre elles est la voiture autonome, pour lequel de nombreux efforts ont été déployés au cours du siècle dernier pour le rendre réel. L'évaluation des algorithmes proposés constitue aujourd'hui un goulot d'étranglement très important pour la voiture autonome. En raison du grand nombre de scénarios possibles, il n'est pas possible de l'effectuer dans la vie réelle. Une alternative consiste à simuler des environnements virtuels où toutes les configurations possibles peuvent être établies à l'avance. Cependant, ces environnements ne sont pas aussi réalistes que le monde réel. Dans cette thèse, nous avons étudié la pertinence d'inclure des images hyperspectrales dans la création de nouveaux environnements virtuels. De plus, nous avons proposé de nouvelles méthodes pour améliorer la compréhension des scènes 3D pour la voiture autonome.

In the field of composite materials, mesoscale modeling based on X-ray computed tomography are very widespread nowadays. This descriptive method requires image processing to identify the different objects within the material. In the present study, two different instance segmentation approaches are proposed: 1) a method based on Mathematical Morphology and 2) a Deep Learning one. Both methods are applied to determine the yarns paths and their envelopes. We succeed in both tasks on a dry 3D ply-to-ply angle-interlock fabric at low compaction level. In absence of manual labelling of the yarns envelopes, we manage to train a Deep Convolutional Neural Network (DCNN) on the pseudo-labeling provided by the morphological method, improving the latter and showing the potential of deep learning for image segmentation in this context, when yarns cross-sections are labelled by distance functions. At a higher level of compaction, we also manage to recover the yarn paths thanks to deep learning.

Empirical mode decomposition (EMD) is a fully data driven method for multiscale decomposing signals into a set of components known as intrinsic mode functions. EMD is based on lower and upper envelopes of the signal in an iterated decomposition scheme. In this paper, we put forward a simple yet effective method to learn EMD from data by means of morphological operators. We propose an end-to-end framework by incorporating morphological EMD operators into deeply learned representations, trained using standard backpropagation principle and gradient descent-based optimization algorithms. Three generalizations of morphological EMD are proposed: a) by varying the family of structuring functions, b) by varying the pair of morphological operators used to calculate the envelopes, and c) the use of a convex sum of envelopes instead of the classical mean point used in classical EMD. We discuss in particular the invariances that are induced by the morphological EMD representation. Experimental results on supervised classification of hyperspectral images by 1D convolutional networks demonstrate the interest of our method.

International audience; Hierarchical clustering (HC) is a powerful tool in data analysis since it allows discovering patterns in the observed data at different scales. Similarity-based HC methods take as input a fixed number of points and the matrix of pairwise similarities and output the dendrogram representing the nested partition. However, in some cases, the entire dataset cannot be known in advance and thus neither the relations between the points. In this paper, we consider the case in which we have a collection of realizations of a random distribution, and we want to extract a hierarchical clustering for each sample. The number of elements varies at each draw. Based on a continuous relaxation of Dasgupta's cost function, we propose to integrate a triplet loss function to Chami's formulation in order to learn an optimal similarity function between the points to use to compute the optimal hierarchy. Two architectures are tested on four datasets as approximators of the similarity function. The results obtained are promising and the proposed method showed in many cases good robustness to noise and higher adaptability to different datasets compared with the classical approaches.

International audience; For the development of a new porous material such as catalytic carrier, the control of the textural properties is of fundamental importance. In order to move towards rational synthesis, it is necessary to better understand the physical phenomena that generate a defined solid structure. A contribute to this purpose can be achieved by studying the aggregation process inside colloidal suspensions, leading to porosity generation: this phenomenon can be described with a Brownian dynamics model that, for any set of chemical parameters, gives access to the mass distribution and the fractal dimension of colloidal aggregates. However, this model cannot be used for the simulation of large colloidal systems, due to its high computational time, limiting comparison with analytical methods, which probe the whole multi-scale system. This problem is solved by developing a new aggregation morphological model, wherein the fractal dimension is tuned with two compactness parameters. An efficient simulation algorithm is proposed in case of spheres, for which the fractal dimension of the generated aggregates varies between 1.2 and 3. Brownian dynamics results are used to parametrize this purely geometric model, in order to constrain the size and the morphology of the aggregates created. The large numerical solid will be representative of the textural properties of a real solid and will give more information on the porous network. It could be used, for example, to simulate diffusive transport coupled with chemical reaction and to study the impact of the geometry of the porous system on the catalytic performance.

International audience; Geometric tortuosity is one of the foremost topological characteristics of porous media. Despite the various definitions in the literature, to our knowledge, they are all linked to an arbitrary propagation direction. This paper proposes a novel topological descriptor, named M-tortuosity, by giving a more straightforward definition, describing the data regardless of physicochemical processes. M-tortuosity, based on the concept of geometric tortuosity, is a scalable descriptor, meaning that information of several dimensions (scalar, histograms, 3D maps) is available. It is applicable on complex disconnected structures without any arbitrary definition of entry and exit. Topological information can be represented by aggregation into a unique scalar descriptor for classification purposes. It is extended by iterative erosions to take into account porous structure narrowness, especially bottleneck effects. This new descriptor, called M-tortuosity-by-iterative-erosions, describes tortuosity of the porous part as seen by a spherical particle of given size walking along the network. Boolean models are used to simulate different porous media structures in order to test the proposed characterization.

International audience; The electronic conductivity, at the multiscale, of lithium-ion positive composite electrodes based on LiNi_{1/3}Mn_{1/3}Co_{1/3}O_2 and/or carbon-coated LiFePO_4, carbon black and poly(vinylidene fluoride) mixture is modeled. The electrode microstructures are acquired numerically in 3D by X-ray tomography and FIB/SEM nanotomography and numerically segmented to perform electrostatic simulations using Fast Fourier Transform (FFT) method. Such simulations are easy and quick to perform because they are directly computed on the grid represented by the voxels in the 3D volumes. Numerical results are compared with experimental measurements of the multiscale electronic conductivity by broadband dielectric spectroscopy (BDS). A good prediction is realized for the bulk conductivities of the C/LiFePO_4 phase and the CB/PVdF mixture. The combination of X-ray and FIB/SEM tomography, FFT simulation method, and BDS is thus well adapted to understand the influence of the electrode composition and microstructure on its electronic conductivity.

International audience

Random projection is a common technique for designing algorithms in a variety of areas, including information retrieval, compressive sensing and measuring of outlyingness. In this work, the original random projection outlyingness measure is modified and associated with a neural network to obtain an unsupervised anomaly detection method able to handle multimodal normality. Theoretical and experimental arguments are presented to justify the choices of the anomaly score estimator, the dimensions of the random projections, and the number of such projections. The contribution of adapted dropouts is investigated, along with the affine stability of the proposed method. The performance of the proposed neural network approach is comparable to a state-of-the-art anomaly detection method. Experiments conducted on the MNIST, Fashion-MNIST and CIFAR-10 datasets show the relevance of the proposed approach, and suggest a possible extension to a semi-supervised setup.

International audience; An adaptive three-dimensional Kalman filter for the tracking of maneuvering targets in three dimensions is proposed. In the radar industry, numerous trackers are based on a constant velocity model, with a process noise covariance matrix Q which is adapted in real time to enhance tracking: it is kept at moderate values during straight lines where the constant velocity assumption applies and is increased during maneuvers. In the present paper we advocate a novel method to increase Q during maneuvers (and hence the Kalman gains) based on a recurrent neural network (RNN). The difficulty and the interest of our approach lies in the fact the neural network is trained together with the filter, by backpropagation through the filter, and hence learns the covariance matrix such as to directly maximize the accuracy of the final output.

International audience; The translation equivariance of convolutions can make convolutional neural networks translation equivariant or invariant. Equivariance to other transformations (e.g. rotations, affine transformations, scalings) may also be desirable as soon as we know a priori that transformed versions of the same objects appear in the data. The semigroup cross-correlation, which is a linear operator equivariant to semigroup actions, was recently proposed and applied in conjunction with the Gaussian scale-space to create architectures which are equivariant to discrete scalings. In this paper, a generalization using a broad class of liftings, including morphological scale-spaces, is proposed. The architectures obtained from different scale-spaces are tested and compared in supervised classification and semantic segmentation tasks where objects in test images appear at different scales compared to training images. In both classification and segmentation tasks, the scale-equivariant architectures improve dramatically the generalization to unseen scales compared to a convolutional baseline. Besides, in our experiments morphological scale-spaces outperformed the Gaussian scale-space in geometrical tasks.

International audience; Mathematical morphology is a useful theory of nonlinear operators widely used for image processing and analysis. Despite the successful application of morphological operators for binary and gray-scale images, extending them to vector-valued images is not straightforward because there are no unambiguous orderings for vectors. Among the many approaches to multivalued mathematical morphology, those based on total orders are particularly promising. Morphological operators based on total orders do not produce the so-called false-colors. On the downside, they often introduce irregularities in the output image. Although the irregularity issue has a rigorous mathematical formulation, we are not aware of an efficient method to quantify it. In this paper, we propose to quantify the irregularity of a vector-valued morphological operator using the Wasserstein metric. The Wasserstein metric yields the minimal transport cost for transforming the input into the output image. We illustrate by examples how to quantify the irregularity of vector-valued morphological operators using the Wasserstein metric.

Microscopy-image analysis provides precious information on size and structure of colloidal aggregates and agglomerates. The structure of colloids is often characterized using the three-dimensional mass fractal dimension d3Df, which is different from the two dimensional fractal dimension d2Dfthat can be computed from microscopy-images. In this work we propose to use a recent morphological aggregation model to find a relationship between 2D image fractal dimension and 3D mass fractal dimension of aggregates and agglomerates. Our case study is represented by scanning transmission electron microscopy-images of boehmite colloidal suspensions. The behaviour of the computed d3Df at different acid and base concentration shows a fair agreement with the results of Small Angle X-Ray Scattering and with the literature, enabling to use the d3Df vs d2Df relationship to study the impact of the composition of the colloidal suspension on the density of colloidal aggregates and agglomerates .

When too few field measurements are available for the geological mod-ling of complex folded structures, results of implicit methods typically exhibit an unsatisfactory bubbly aspect. However, in such cases, anisotropy data are often readily available but not fully exploited. Among them, fold axis data are a straightforward indicator of this local anisotropy direction. Focusing on the so-called potential field method, this work aims to evaluate the effect of the incorporation of such data into the modeling process. Given locally sampled fold axis data, this paper proposes to use the second-order derivatives of the scalar field in addition to the existing first-order ones. The mathematical foundation of the approach is developed and the respective efficiencies of both kinds of constraints are tested. Their integration and impact are discussed based on a complex case study, thereby providing practical guidelines to geomodeling tool users on the parsimonious use of data for the geological modeling of complex folded structures

During recent years, the renaissance of neural networks as the major machine learning paradigm and more specifically, the confirmation that deep learning techniques provide state-of-the-art results for most of computer vision tasks has been shaking up traditional research in image processing. The same can be said for research in communities working on applied harmonic analysis, information geometry, variational methods, etc. For many researchers, this is viewed as an existential threat. On the one hand, research funding agencies privilege mainstream approaches especially when these are unquestionably suitable for solving real problems and for making progress on artificial intelligence. On the other hand, successful publishing of research in our communities is becoming almost exclusively based on a quantitative improvement of the accuracy of any benchmark task. As most of my colleagues sharing this research field, I am confronted with the dilemma of continuing to invest my time and intellectual effort on mathematical morphology as my driving force for research, or simply focussing on how to use deep learning and contributing to it. The solution is not obvious to any of us since our research is not fundamental, it is just oriented to solve challenging problems, which can be more or less theoretical. Certainly, it would be foolish for anyone to claim that deep learning is insignificant or to think that one's favourite image processing domain is productive enough to ignore the state-of-the-art. I fully understand that the labs and leading people in image processing communities have been shifting their research to almost exclusively focus on deep learning techniques. My own position is different: I do think there is room for progress on mathematically grounded image processing branches, under the condition that these are rethought in a broader sense from the deep learning paradigm. Indeed, I firmly believe that the convergence between mathematical morphology and the computation methods which gravitate around deep learning (fully connected networks, convolutional neural networks, residual neural networks, recurrent neural networks, etc.) is worthwhile. The goal of this talk is to discuss my personal vision regarding these potential interactions. Without any pretension of being exhaustive, I want to address it with a series of open questions, covering a wide range of specificities of morphological operators and representations, which could be tackled and revisited under the paradigm of deep learning. An expected benefit of such convergence between morphology and deep learning is a cross-fertilization of concepts and techniques between both fields. In addition, I think the future answer to some of these questions can provide some insight on understanding, interpreting and simplifying deep learning networks.

This paper deals with the relationship between spectral analysis in min-max algebra and ultrametric morphological operators. Indeed, morphological semigroups in ultrametric spaces are essentially based on that algebra. Theory of eigenfunctionals in min-max analysis is revisited, including classical applications (preference analysis, percolation and hierarchical segmentation). Ultrametric distance is the fix point functional in min-max analysis and from this result, we prove that the ultrametric distance is the key ingredient to easily define the eigenfunctions of ultrametric morphological openings and closings.

Shock sensitivity of explosive formulations is linked to their microstructure. Shock waves interact with the heterogeneitiesof the microstructure which lead to localized increase of pressure and temperature called hot spots. The hot spots can potentially initiate the detonation. The objective of the PhD thesis is to study the influence of different microstructural parameters on the sensitivity to shock of energetic materials. First, the micro-computed tomography images of three RDX/Wax materials (70/30 in weight) are studied. These three materials have the same composition but different microstructures and different sensitivity to shock. The micro-computed tomography images are filtered, segmented and characterized. The characterization allows the identification and quantification of the differences between the three materials: granulometry, spatial distribution, grains shape, contact points between grains and elastic response to deformation. Then, each labeled grain is extracted and characterized with several descriptors (volume, surface, mean radius, sphericity, angularity) to create a library of grains. The results of image analysis are compared with experimental data to evaluate the effectiveness of the segmentation tool. Then, virtual microstructures are generated with the library of grains. It allows to control the characteristics of the grains implanted in the virtual material and to isolate the effects of some microstructural parameters, while respecting the particle size and spatial distribution of the real material. Dynamic numerical simulations have demonstrated the importance of the intra-granular defects and of the contact points between the grains in the localization of stresses and the formation of hotspots. The role of the contact points between the grains was already highlighted in FFT elastic quasi-static simulations. In addition to this numerical work, shock recovery experiments were developed and performed at the ISL. Damaged RDX/Wax samples were imaged with micro-computed tomography. The first visual inspections reveal numerous cracks in the grains and an important grain-wax decohesion for shock pressure largely under the detonation threshold.; La sensibilité au choc d’un matériau énergétique est liée à sa microstructure. L’onde de choc interagit avec la microstructure en créant des sauts de pression et des hausses de température localisés que l’on appelle points chauds. Ces points chauds sont de potentiels sites d’initiation de la détonation. L’objectif de thèse est d’étudier l’influence de différents paramètres microstructuraux sur la sensibilité au choc d’un matériau énergétique. Dans un premier temps, les microtomographies de trois matériaux RDX/Cire (70/30 en masse) sont étudiées. Ces trois matériaux sont identiques en termes de composition mais diffèrent par leur microstructure ainsi que par leur seuil d’amorçage en détonation. Les microtomographies sont filtrées et segmentées avant d’être caractérisées. La caractérisation permet d’identifier et de quantifier les différences de granulométries, distributions spatiale, formes des grains, points de contact entre les grains et réponses élastiques des 3 matériaux. Chaque grain labellisé est ensuite étudié individuellement (volume, surface, rayon moyen, sphéricité, convexité, angularité) et extrait pour former une bibliothèque de grains. Les résultats d’analyse d’images sont comparés avec des mesures expérimentales pour vérifier la validité de l’outil de segmentation mis en œuvre. Dans un second temps, des microstructures virtuelles sont générées à partir de la bibliothèque de grains réels afin de contrôler les caractéristiques des grains présents dans le matériau virtuel et d’isoler les effets de différents paramètres microstructuraux, tout en respectant la granulométrie et la distribution spatiale du matériau réel. Des simulations numériques dynamiques ont mis en évidence le rôle des défauts intra-granulaires ainsi que des points de contact entre les grains dans la localisation des contraintes et la formation de points chauds. L’importance des points de contact entre les grains était déjà pressentie à la suite de calculs FFT élastique quasi statique. Parallèlement à ce travail numérique, des expériences de récupération sous choc ont été mises au point et réalisées à l’ISL. Les échantillons endommagés sont microtomographiés. Les premières observations visuelles révèlent d’importantes fissurations et décohésions pour des pressions de choc bien inférieures au seuil d’amorçage de la détonation

This work presents a study in non-linear mechanics, in the classical framework of brittle fracture in heterogeneous media. We study numerically the localization and propagation of damage in a triamino-trinitrobenzene-based (TATB) material under various mechanical and thermal loadings. The polycrystalline microstructure, which in addition contains a binder and porosities, is partially characterized by electron microscopy images as well as experimental testings under mechanical and thermal solicitations, and presents in its initial state, residual stresses and, possibly, pre-existing cracks. We first develop a numerical method based on fast Fourier transforms for predicting the material's brittle response. A "phase field" is used to describe local damage within the microstructure. The proposed algorithms, validated by comparisons with data obtained in the literature, takes into account the irreversible nature of damage as well as the strong crystal anisotropy in TATB grains, in elasticity and tenacity. Secondly, we apply the developed numerical schemes to the study of TATB polycrystals. We focus first on transgranular cracking under a cooling-heating thermal cycle, after which the role of the porosity and of the binder as well as purely-mechanical loadings are investigated. These various phenomena are taken into account incrementally and identified by an inverse method using available experimental data. The results obtained in the present work show the preponderant role played by inter and transgranular cracks on the thermal behavior under thermal and mechanical loadings. Moreover our approach also quantifies the relative influence of various mechanisms : porosity, binder and grains elasticity, but also the grains mechanical anisotropy and pre-existing cracks in the material in its initial state.; Ce travail s'inscrit dans la thématique classique en mécanique non-linéaire de la modélisation du comportement à rupture de milieux hétérogènes. On étudie par des moyens numériques la localisation et la propagation de l'endommagement dans un comprimé au TATB (triamino-trinitrobenzène) soumis à diverses sollicitations mécaniques et thermiques. La microstructure polycristalline, qui contient en outre liant et porosités, est partiellement caractérisée à travers des images obtenues par microscopie électronique, ainsi que des essais mécaniques et thermiques, et présente, à l'état initial, des contraintes résiduelles, voir des fissures pré-existantes. On développe dans un premier temps une méthode numérique de calculs par transformées de Fourier rapide pour la prédiction de l'endommagement reposant sur l'utilisation d'un "champ de phase", qui décrit l’endommagement local au sein de la microstructure. Les algorithmes proposés, validés à partir de données issues de la littérature, prennent en compte le caractère irréversible de l'endommagement ainsi que la forte anisotropie des cristaux de TATB, en élasticité maiségalement du point de vue de la fissuration. Dans un second temps, on applique les schémas numériques développés aupolycristal étudié. On s’intéresse en particulier à la fissuration transgranulaire sous cycle thermique à froid, puis on intègre la porosité et la fissuration dans le liant sous chargement thermique et en traction. Ces divers phénomènes sont pris en compte de manière incrémentale et identifiés par méthode inverse sur les données expérimentales disponibles. Les résultats de ce travail montrent d'une part le rôle prépondérant joué par l'endommagement inter et transgranulaire sur le comportement thermique à froid et lors des essais mécaniques. La démarche entreprise permet, en outre, de préciser l'influence relative des divers mécanismes pris en compte : porosité, élasticité dans le liant et dans les grains, mais également anisotropie mécanique des grains ou endommagement du matériau à l'état initial.

Graphs are powerful mathematical structures representing a set of objects and the underlying links between pairs of objects somehow related. They are becoming increasingly popular in data science in general and in particular in image or 3D point cloud analysis. Among the wide spectra of applications, they are involved in most of the hierarchical approaches.Hierarchies are particularly important because they allow us to efficiently organize the information required and to analyze the problems at different levels of detail. In this thesis, we address the following topics. Many morphological hierarchical approaches rely on the Minimum Spanning Tree (MST). We propose an algorithm for MST computation in streaming based on a graph decomposition strategy. Thanks to this decomposition, larger images can be processed or can benefit from partial reliable information while the whole image is not completely available.Recent LiDAR developments are able to acquire large-scale and precise 3D point clouds. Many applications, such as infrastructure monitoring, urban planning, autonomous driving, precision forestry, environmental assessment, archaeological discoveries, to cite a few, are under development nowadays. We introduce a ground detection algorithm and compare it with the state of the art. The impact of reducing the point cloud density with low-cost scanners is studied, in the context of an autonomous driving application. Finally, in many hierarchical methods similarities between points are given as input. However, the metric used to compute similarities influences the quality of the final results. We exploit metric learning as a complementary tool that helps to improve the quality of hierarchies. We demonstrate the capabilities of these methods in two contexts. The first one,a texture classification of 3D surfaces. Our approach ranked second in a task organized by SHREC’20 international challenge. The second one learning the similarity function together with the optimal hierarchical clustering, in a continuous feature-based hierarchical clustering formulation.; Les graphes sont de puissantes structures mathématiques représentant un ensemble d’objets et les relations sous-jacentes entre eux. Ils sont de plus en plus populaires, en particulier dans l’analyse hiérarchique des images ou des nuages de points 3D. Les hiérarchies sont très répandues car elles nous permettent d’organiser efficacement l’information et d’analyser les problèmes à différents niveaux de détail. Dans cette thèse, nous abordons les sujets suivants : De nombreuses approches hiérarchiques morphologiques s’appuient sur l’arbre de poids minimum (MST). Nous proposons un algorithme pour le calcul du MST en streaming reposant sur une stratégie de décomposition des graphes. Grâce à cette décomposition, des images plus grandes peuvent être traitées ou peuvent bénéficier d’une information partielle fiable alors que l’image entière n’est pas encore disponible. Les récents développements du lidar permettent d’acquérir des nuages de points 3D précis et à grande échelle. De nombreuses applications, telles que la surveillance des infrastructures, l’urbanisme, la conduite autonome, l’agriculture de précision, pour n’en citer que quelques-unes, sont en cours de développement. Nous introduisons un algorithme de détection du sol et le comparons à l’état de l’art. L’impact de la réduction de la densité des nuages de points avec des scanners à faible coût est étudié. Enfin, dans de nombreuses méthodes hiérarchiques, les similarités entre les points sont données en entrée. Cependant, la métrique utilisée pour calculer les similitudes influence la qualité des résultats. Nous abordons l’apprentissage de la métrique comme un outil complémentaire qui contribue à améliorer la qualité des hiérarchies. Nous démontrons les capacités de ces méthodes dans deux contextes. Le premier, une classification de la texture des surfaces 3D, classée deuxième dans une tâche du défi international SHREC’20. Le second permet d’apprendre la fonction de similarité ainsi que la hiérarchie optimale, dans une formulation continue dans l’espace des caractéristiques

International audience

The works deals with a macroscopically isotropic energetic material based on triamino-trinitrobenzene (TATB) crystals bonded with a small volume fraction of a thermoplastic polymer. This material is shown experimentally to display an irreversible thermal expansion behavior characterized by dilatancy and variations of its thermal expansion coefficient when heated or cooled outside a narrow reversibility temperature range. The analysis of cooling results suggests the existence of residual stresses in the initial state, attributed to the manufacturing process. Microstructure-level FFT computations including the very strong anisotropic thermoelastic TATB crystal response and temperature-dependent binder plasticity, show that strong internal stresses develop in the disoriented crystals under thermal load, either heating or cooling. Upon cooling, binder plastic yielding in hindered, thus promoting essentially brittle microcracking, while it is favored upon heating. Despite its low volume fraction, the role of the binder is essential, its plastic yielding causing stress redistribution and residual stresses upon cooling back to ambient, .

In this work, we are interested in the study of the local and global regularity of a class of functions which are relevant in fractal analysis, the so-called Hölder continuous functions. Indeed, fractal dimension and Hölder exponent of functions are related in many cases. Estimates of the dimension or the exponent of this kind of functions are classicaly based either on wavelet theory or on multiscale morphological operators. In this paper, Hölder function characterization is revisited from the mathematical morphology viewpoint, including the connection with some contributions from the field of max-plus mathematics. We show in particular that morphological operators on metric spaces are naturally formulated in the case of equicontinuous functions, including Hölder functions. We focus on the case of morphological semigroups on length spaces since they provide the natural extension of multiscale morphological operators on the Euclidean space. We prove how these semigroups can be used to characterize the exponent of Hölder functions on length spaces.

Convolutional Neural Network (CNNs) models&rsquo, size reduction has recently gained interest due to several advantages: energy cost reduction, embedded devices, and multi-core interfaces. One possible way to achieve model reduction is the usage of Rotation-invariant Convolutional Neural Networks because of the possibility of avoiding data augmentation techniques. In this work, we present the next step to obtain a general solution to endowing CNN architectures with the capability of classifying rotated objects and predicting the rotation angle without data-augmentation techniques. The principle consists of the concatenation of a representation mapping transforming rotation to translation and a shared weights predictor. This solution has the advantage of admitting different combinations of various basic, existing blocks. We present results obtained using a Gabor-filter bank and a ResNet feature backbone compared to previous other solutions. We also present the possibility to select between parallelizing the network in several threads for energy-aware High Performance Computing (HPC) applications or reducing the memory footprint for embedded systems. We obtain a competitive error rate on classifying rotated MNIST and outperform existing state-of-the-art results on CIFAR-10 when trained on up-right examples and validated on random orientations.

International audience; In this article an “hyper-reduced” scheme for the Crisfield’s algorithm (Crisfield, 1981) applied to buckling simulations and plastic instabilities is presented. The two linear systems and the ellipse equation entering the algorithm are projected on a reduced space and solved in a reduced integration domain, resulting in a system of “hyper-reduced” equations. Use is made of the Gappy proper orthogonal decomposition to recover stresses outside the reduced integration domain. Various methods are proposed to construct a reduced bases, making use of simulation data obtained with standard finite element method and a stress-based error criterion for the hyper reduced calculations is proposed. A “greedy” algorithm coupled with this error criterion is used to generate intelligently full standard finite element simulations and enrich the reduced base, demonstrating the adequacy of the error criterion. Finally, numerical results pertaining to elastoplastic structures undergoing finite strains, with emphasis on buckling and limit load predictions are presented. A parametric study on the geometry of the structure is carried out in order to determine the domain of validity of the proposed hyper-reduced modeling approach.

We develop a methodology for classifying defects based on their morphology and induced mechanical response. The proposed approach is fairly general and relies on morphological operators (Angulo and Meyer in 9th international symposium on mathematical morphology and its applications to signal and image processing, pp. 226-237, 2009) and spherical harmonic decomposition as a way to characterize the geometry of the pores, and on the Grassman distance evaluated on FFT-based computations (Willot in C. R., Méc. 343(3):232–245, 2015), for the predicted elastic response. We implement and detail our approach on a set of trapped gas pores observed in X-ray tomography of welded joints, that significantly alter the mechanical reliability of these materials (Lacourt et al. in Int. J. Numer. Methods Eng. 121(11):2581–2599, 2020). The space of morphological and mechanical responses is first partitioned into clusters using the “k-medoids” criterion and associated distance functions. Second, we use multiple-layer perceptron neural networks to associate a defect and corresponding morphological representation to its mechanical response. It is found that the method provides accurate mechanical predictions if the training data contains a sufficient number of defects representing each mechanical class. To do so, we supplement the original set of defects by data augmentation techniques. Artificially-generated pore shapes are obtained using the spherical harmonic decomposition and a singular value decomposition performed on the pores signed distance transform. We discuss possible applications of the present method, and how medoids and their associated mechanical response may be used to provide a natural basis for reduced-order models and hyper-reduction techniques, in which the mechanical effects of defects and structures are decorrelated (Ryckelynck et al. in C. R., Méc. 348(10–11):911–935, 2020).

International audience; Cette étude présente les outils permettant, d’une part, de caractériser les microstructures hétérogènes et leur morphologie et, d’autre part, de les modéliser, à l’aide d’ensembles aléatoires. Elle aborde les ensembles de points aléatoires de Poisson, les modèles booléens, les modèles à sphères dures et quelques modèles de partitions aléatoires et de champs gaussiens. Plusieurs exemples d’application sont présentés.

International audience; This book covers methods of Mathematical Morphology to model and simulate random sets and functions (scalar and multivariate). The introduced models concern many physical situations in heterogeneous media, where a probabilistic approach is required, like fracture statistics of materials, scaling up of permeability in porous media, electron microscopy images (including multispectral images), rough surfaces, multi-component composites, biological tissues, textures for image coding and synthesis. The common feature of these random structures is their domain of definition in n dimensions, requiring more general models than standard Stochastic Processes.The main topics of the book cover an introduction to the theory of random sets, random space tessellations, Boolean random sets and functions, space-time random sets and functions (Dead Leaves, Sequential Alternate models, Reaction-Diffusion), prediction of effective properties of random media, and probabilistic fracture theories.Most basic models of random structures are thoroughly investigated and illustrated, starting from their mathematical construction, their probabilistic properties, and their identification from data

International audience; This paper aims at providing an overview of the use of mathematical morphology, in its algebraic setting, in several fields of artificial intelligence (AI). Three domains of AI will be covered. In the first domain, mathematical morphology operators will be expressed in some logics (propositional, modal, description logics) to answer typical questions in knowledge representation and reasoning, such as revision, fusion, explanatory relations, satisfying usual postulates. In the second domain, spatial reasoning will benefit from spatial relations modeled using fuzzy sets and morphological operators, with applications in model-based image understanding. In the third domain, interactions between mathematical morphology and deep learning will be detailed. Morphological neural networks were introduced as an alternative to classical architectures, yielding a new geometry in decision surfaces. Deep networks were also trained to learn morphological operators and pipelines, and morphological algorithms were used as companion tools to machine learning, for pre/post processing or even regularization purposes. These ideas have known a large resurgence in the last few years and new ones are emerging.