<\/p>\n\n\n\n
H\u00f4tel Dupanloup<\/a> <\/strong>au centre ville d’Orl\u00e9ans (Google maps<\/a>)<\/p>\n\n\n\n <\/p>\n","protected":false},"excerpt":{"rendered":" Lieu H\u00f4tel Dupanloup au centre ville d’Orl\u00e9ans (Google maps) Programme Jeudi 23 juin 2022 14h-17h30 : Expos\u00e9s scientifiques de chercheurs invit\u00e9s Programme : Accueil – caf\u00e9 14h30-15h15 : Hermine Chatoux, The metrological use of fractals to study human perception and physical sensors 15h15-16h … Continuer la lecture Programme<\/h1>\n\n\n\n
Jeudi 23 juin 2022<\/h2>\n\n\n\n
Programme :<\/strong>Vendredi 24 juin 2022<\/h2>\n\n\n\n
Programme :<\/strong>H\u00f4tels recommand\u00e9s<\/h2>\n\n\n\n
R\u00e9sum\u00e9s des expos\u00e9s<\/h2>\n\n\n\n
Despite their impressive results in most image restoration tasks, deep learning approaches fail to generalise to the slightest changes in either the acquisition modality or the distortion modality. To that end, the most common solution is to acquire a new database for each modality and retrain a new neural network, which is both costly and time consuming. To circumvent these limitations, we propose to train deep learning restoration methods on images generated with a simple dead leaves model equipped with a carefully designed distortion model. In this talk, I will first introduce this synthetic image model and the implementation choices we made. Then, I will explain how we modelled image distortions for various image restoration tasks. I will then present some experimental results confirming our intuition that synthetic images and a good deterioration model are sufficient to train a neural network with performances on par with its classical counterparts. Finally, we will give some perspectives to this work, with an ongoing study on the understanding of the distribution of colors in natural images, which is crucial to color dead leaves images.
<\/li>
The RGB color model is an additive color model in which the red, green and blue colors are added to reproduce colors. We consider the color monogenic framework to define random fields according to natural colors. This approach is based on the monogenic representation of a greyscale image, based on Riesz transform and particularly well-adapted to detect directionality of self-similar of operator scaling Gaussian fields. This talk is dedicated to the memory of Mark Marvin Meerschaert (1955-2020).
<\/li>
To study the human perception of texture, the fractional Brownian movement can be used to generate color texture images (a controlled variation in the spatio-chromatic space). Indeed, fractal textures are not directly related to a semantic subject and therefore will allow to measure \u201ceye sensation\u201d and not \u201cbrain analysis\u201d. Our focus of interest is the physical and perceptual constraints to generate such a mathematical object. The fBm strong hypothesis is the continuity of the space. When applying the concept to color images, a quantification is realized both in the spatial and the chromatic space.
<\/li>
Suppose that a space is divided into two regions that are separated by a locally smooth boundary. Superpose the space with a regular lattice structure, and for each vertex of the lattice, record which region it belongs to. From the samples at the lattice vertices, our goal is to estimate the measure of the boundary. In this talk, I describe an algorithm that recovers the exact measure of the boundary as the size of the lattice cell approaches zero. This algorithm can be used to estimate the perimeter of excursion sets of 2D, anisotropic random fields, for example, with desirable probabilistic properties (see our preprint hal-03582844<\/a>).
<\/li>
De nombreux r\u00e9sultats concernant les propri\u00e9t\u00e9s g\u00e9om\u00e9triques des champs al\u00e9atoires ont \u00e9t\u00e9 \u00e9tablis dans le cas des champs gaussiens ; citons par exemple la longueur des lignes nodales ou la caract\u00e9ristique d\u2019Euler qui d\u00e9crivent la structure topologique des excursions. Par ailleurs, les mod\u00e8les les plus utilis\u00e9s dans le cadre de l\u2019analyse statistique des valeurs extr\u00eames de processus spatiaux sont les processus de m\u00e9lange gaussien, en \u00e9chelle ou en position. Dans cet expos\u00e9, on commencera par pr\u00e9senter quelques r\u00e9sultats concernant la g\u00e9om\u00e9trie des excursions gaussiennes et on montrera comment les utiliser en statistique inf\u00e9rentielle pour exhiber des propri\u00e9t\u00e9s du champ consid\u00e9r\u00e9. Ensuite, on \u00e9tendra ces r\u00e9sultats \u00e0 certains mod\u00e8les de m\u00e9lange ainsi qu\u2019\u00e0 des processus limites connus sous le nom de processus de Pareto g\u00e9n\u00e9ralis\u00e9s.
Travail en collaboration avec Elena Di Bernardino (LJAD, Nice) et Thomas Opitz (BioSP, INRAE Avignon).<\/li><\/ul>\n\n\n\n
Le bruit proc\u00e9dural est utilis\u00e9 dans les applications de rendu en temps r\u00e9el pour texturer des surfaces de taille non born\u00e9e afin d’\u00e9viter les artefacts de r\u00e9p\u00e9tition et d’alignement d’un pavage p\u00e9riodique classique. Comme les bruits proc\u00e9duraux impl\u00e9mentent g\u00e9n\u00e9ralement des champs stationnaires Gaussiens, le type de textures pouvant \u00eatre g\u00e9n\u00e9r\u00e9 est statistiquement limit\u00e9. Nous avons montr\u00e9 l’ann\u00e9e derni\u00e8re qu’il \u00e9tait possible d’instancier des champs cyclostationnaires Gaussiens pour g\u00e9n\u00e9rer en plus des textures dont l’organisation globale est p\u00e9riodique, soit \u00e0 partir de spectres, soit \u00e0 partir d’un exemple d’entr\u00e9e, en adaptant certains des algorithmes existants. Cette pr\u00e9sentation va principalement concerner notre article de Eurographics 2021 (https:\/\/hal.archives-ouvertes.fr\/hal-03181139\/<\/a>), de quelques d\u00e9veloppements th\u00e9oriques plus r\u00e9cents, et de pistes de recherches.<\/li><\/ul>\n\n\n\n