Centralesupelec

Ecole Centrale De Paris

Ecole Centrale Paris

Therapanacea

Nikos Paragios

Stavros Alchatzidis

Nikos Komodakis

Benjamin Glocker

Kumar Shreshtha

Georgios Tziritas

Aristeidis Sotiras

Enzo Ferrante

Aurelien Lombard

Singht Bharat

Rafael Marini Silva

Images are synthesized from a source to a target nature through unsupervised machine learning (ML), based on an original training set of unaligned source and target images, by training a first ML architecture through an unsupervised first learning pipeline applied to the original set, to generate a first trained model and induced target images consisting in representations of original source images compliant with the target nature. A second ML architecture is trained through a supervised second learning pipeline applied to an induced training set of aligned image pairs, each including first and second items corresponding respectively to an original source image and the induced target image associated with the latter, to generate a second trained model enabling image syntheses from the source to the target nature. Also, applications to effective medical image translations.

Devices and process for synthesizing images from a source nature to a target nature

The invention concerns a method for deformable fusion of a source multi-dimensional image (s(x)) and a target multi-dimensional image (t(x)) of an object, each image being defined on a multi-dimensional domain by a plurality of image signal samples, each sample having an associate position in the multi-dimensional domain and an intensity value, the method comprising estimating a smooth deformation field (d(x)) that optimizes a similarity criterion between the source image and the target image using a Markov Random Field framework, in near real-time performance. The similarity criterion is computed on transform coefficients obtained by applying a sub-space hierarchical transform to the image samples of the target image and to image samples obtained from the source image, an optimal tradeoff between a smoothness condition and the similarity criterion being automatically determined.

Method for real-time deformable fusion of a source multi-dimensional image and a target multi-dimensional image of an object

A method and device for elastic registration of a two-dimensional source digital image of an object of interest with a slice of a three dimensional target volume of the object of interest is provided which defines a Markov Random Field framework having at least one undirected pairwise graph superimposed a two-dimensional image domain having at least a set of regular vertices and at least a set of edges, and defines a grid of control points, each control point corresponding to a vertex of the set of regular vertices, and a neighborhood system of edges associated with vertices. The method also defines a set of multi-dimensional labels of a discrete space to apply a displacement to each control point, the control point displacement defining a transformation adapted to obtain a plane slice of the target volume and an in-plane deformation transformation of the source digital image.

Method and device for elastic registration between a two-dimensional digital image and a slice of a three-dimensional volume with overlapping content

A message passing scheme for MAP inference on Markov Random Fields based on a message computation using an intermediate input vector I, an output message vector M, an auxiliary seed vector S, all of equal length N, and a pairwise function r=d(x,y), where r, x, y are real numbers, includes:

Method and device for efficient parallel message computation for map inference

A method for determining an optimal labeling of pixels in computer vision includes modeling an image by a graph having interior nodes and edges where each image point p is associated with a graph node, each pair of nearest neighbor points p, q is connected by a graph edge, each graph node p is associated with a singleton potential c(p), and each graph edge is associated with a pairwise potential function d(p,q). A label is randomly assigned to each point to initialize unary variables including an indicator function that indicates which label is assigned to which point and dual variables including height variables associated with each node p and label a, and balance variables associated with each edge (p,q) and label a. For each label, a new label c is selected, a capacitated graph is constructed and solved. The label selection divides the image into disjoint regions.

System and method for optimizing single and dynamic markov random fields with primal dual strategies

A method for registering digitized images using Markov Random Fields (MRFs) includes providing a source image f and a target image g, defining a deformation grid of control points, defining a coordinate transformation as where x is a point on the source image, p is a position vector of control point p, dis a displacement vector for each control point, uis a label for point p associated with displacement d, and &#x3b7;( ) is a weighting function for the displacement vector, defining an MRF energy functional to be minimized by T as wherein |G| is a number of control points, |E| is a number of pairs of neighboring control points on a neighborhood system, t is an iteration counter, and associating the MRF with a primary linear program and solving the primary linear program using a fast primal-dual algorithm to yield a coordinate transformation that minimizes the energy functional.

Growing community of inventors

Vincennes, France

Nikos Paragios