Bayesian fusion models based on multi-level constraints and multiple criteria in image processing and computer vision

点击翻译按钮获取中文摘要

My research program investigates the use of new unsupervised (Bayesian) probabilistic or energy-based fusion models for understanding, analyzing, and manipulating still, moving and multidimensional, multispectral or multimodal images.*** More precisely, this research program will attempt to propose new statistical models to synergistically integrate multiple image cues (e.g., color, texture, edges, interest point or symmetry detection, Gestalt perceptual cues, etc.) with possibly different constraints (possibly expressed at different levels of abstraction) in order to better model the intrinsic and complex properties of the (image) solution to be estimated and/or to fuse several weak solutions or different complementary low-level applications (segmentation, edge map, restored image, etc.) in order to achieve either a more reliable and accurate solution or a high-level computer vision task (3D reconstruction, complex shape localization, etc.). *** These models can have a wide range of applications not only in still image processing and computer vision, but also in several other fields, including multi-modal medical image applications, Geoscience imagery and more generally in all multi-camera or multi-modal recognition and reconstruction systems of the next generation.***The adopted framework, for these different research models, mainly relies on the Bayesian statistical theory which allows to take into account some available prior knowledge on the information to be found and to combine this prior model with a (statistical) model describing the interactions between hidden and observed variables (likelihood model). In this framework, the proper use of the available prior information can be expressed by local prior models such as Markov Random Field (MRF) models and contextual knowledge is usually captured through the specification of spatially local interactions or recently through non local (or long-range) interactions. In addition, Bayesian theory makes it also possible to apply global prior (interactions) or constraints such as the natural variability of an object shape to be detected/reconstructed (via global parametric probabilistic prior models or global constraints expressed by the knowledge of a solution at a lower level of abstraction).**