Design of novel algorithms, Models, Texture, Images, Volumes

Parametric approaches for the segmentation and the classification of textures

Regarding image segmentation, we propose an original inverse method for the joint segmentation and deconvolution of textures

This approach is motivated by the assumption that segmentation may be more reliable when performed on a sharper image and that the quality of the deconvolution result may be improved by the knowledge of the regions composing the image. Concerning classification and model selection, we are also working on methods for indirect observation (e.g. including convolution). Proposed solutions are based on the computation of evidences (marginal likelihoods). In both cases, the major difficulty comes from the joint deconvolution. This is addressed within a hierarchical Bayesian framework, estimation/classification tasks being performed by stochastic sampling agorithms.

Image classification and, more specifically texture classification, is also addressed by modeling multiscale spatial dependences and local anisotropy

Current work also concerns the modeling of either time, spectrometric or polarimetric dependences. The multi-scale representation of 2-D or 3-D spatial processes relies on various families of probabilistic models (copulas, generalized multivariate Gaussian densities, SIRV models, etc.). Texture anisotropy is described using the local structure tensor field, itself handled by means of Riemanian probabilistic models of SPD matrices. Many of these theoretical approaches and algorithmic tools have been applied to very high resolution (VHR) remote sensed images for agronomic and environmental purposes : identification and characterization of forest stands, vineyards and oyster racks, etc.

These activities have benefited from the financial support of various entities including the Conseil Regional d’Aquitaine, the french Agence Nationale de la Recherche, the CPU and COTE clusters within Bordeaux University, the Centre National d’Etudes Spatiales (CNES) and the group Safran.

Models and algorithms from information geometry

Many tasks relative to the analysis or processing of 2D or 3D processes are based on the use of descriptors (e.g. statistical moments, parameters of probability distributions, etc.) which may be embedded in spaces which are not consistant with standard tools of Euclidean geometry. Information geometry introduces geometric tools (e.g. distances, centroid, median, separating hyperplane…) into such descriptor spaces which allow to represent and to handle various mathematical objects (e.g. copulas, generalized multivariate gaussian densities, SIRV models, structure tensorts, etc.) useful to the analysis of spatial processes. During the last years, we’ve been developping new models and algorithms based on information geometry, that allow robust and efficient approaches to learning paradigms (classification, segmentation, etc.) and to the analysis and recontruction of spatial proceses (such as tensor-valued images for instance).

These activities have benefited from the financial support of IdEx Bordeaux and of the CPU cluster.

Synthesis of solid textures

Various works have been carried out on the synthesis of 3D textured images (i.e. solid textures), particularly in the case of images showing anisotropic structures, either fibrous or laminar. Both parametric and non parametric approaches, either deterministic or stochastic, have allowed to generate 3D textures from 3D exemplars and even, under some assumptions, from 2D exemplars while preserving geometric structures and patterns at fine or broader scales.

The outcomes of these works concern mainly material sciences and aim at designing virtual materials. Combined with atomistic simulation techniques applied to carbon materials at the nanoscale, the methods we propose allow to predict the existence of 3D nanostructures non observable by imagery, to describe them and to simulate their behaviour. Studied materials are pyrocarbons for the space and aircraft industry and, more recently, nuclear graphites. At a broader scale, progress has also been made regarding the synthesis of fibrous textures. Applications concern the modeling of woven composite material.

This work has been conducted in collaboration with the LCTS lab (CNRS) and financially supported by the French Agence Nationale de la Recherche (Pyroman project) and Direction Générale de l’Armement.

Detection and reconstruction of objects in textured images and volumes

In the context of computer vision applied to the analysis of real samples or scenes, it is often necessary to identify structured objects within images such as wire-like objects, surfaces or volumes. The detection of such objects may be hindered by the presence of noise and acquisition artifacts. In the case of multiple objects, occlusions, junctions or juxtaposition phenomena can also make the objets hardly separable or even detectable.

We propose various approaches dedicated to the detection and the reconstruction of objects in textured images and volumes. In particular we are interested in wire-like structures (in 2D or 3D) and in surfaces (planar or curved). These approaches take advantage of the local anisotropy of textures, described by mathematical tools inherited from differential geometry such as the structure tensor. For instance, the tracking of wire-like structures is based on the 2D/3D orientation flow whereas the identification of surfaces is done using the normal vector field.

Various applications can be found within our industrial or academic partnerships. They include fault detection and horizon reconstruction in seismic images, segmentation of fibrous structures in tomographic images of woven (Carbon-Carbon or Carbon-Ceramic composites) or non woven (bio-based or wood-textile composites) materials. These activities have received financial support from industrial partners (Total, Safran) and public organizations (Direction Générale de l’Armement, ADEME).

MOTIVE's Skills

Classification, segmentation and inverse problems

Characterization of 2D non-gaussian correlated spatial processes

Machine learning / Deep Learning

Analysis of 2D and 3D textures

Computer Vision

Remote and Proximal Sensing

MOTIVE

Parametric approaches for the segmentation and the classification of textures

Models and algorithms from information geometry

Synthesis of solid textures

Detection and reconstruction of objects in textured images and volumes

MOTIVE's Skills

Latest news from the team

Staff

Contact our team