Visual processing

Problematic:

The general purpose is to solve ``classical'' problems of in image processing in the particular context of dynamic satellite imagery. Methods focus more particularly on the following items:

• The modeling of structures existing in these images: the notion of ``structure'' is taken in a broad sense, it may correspond to ``objects'', bi or tri-dimensional, connected or not, more or less deformable, but also to vector fields, or to regions that are not specifically attached to an object, such as turbulence areas. With this objective, the team is studying different modeling tools such as level-sets, or methods coming from the theory of complex signals to characterize multi-scale structures (past Thalweg ARC).
• The implementation of these models to matching and tracking problems (matching occurences of a structure at different dates),
• Motion estimation in an image sequence, in the case the sequence displays fluid or even turbulent flows. This subject has been the starting point of different studies for taking into account different motion constraints, also different conservation and regularization equations...This approach goes closer to physics, where the image acquisition process and the dynamics of the observed phenomenon are better taken into consideration (see next subsection for further description).

Objectives:

Short term objectives are related to image modeling adapted to the type of structures encountered in environmental imagery: fluid motion, large deformation, topological change, turbulence. Tools that have been already studied (spline vector fields, characterization of turbulent areas in a fluid, among others) are currently settled in a static context. They must be extended taking into consideration the structures' dynamics and in particular the interactions between the different structures present in a scene. Moreover, re-writing these models in a multi-scale framework is an important research direction, particularly in the context of satellite imagery, in which one is confronted with the analysis of data having various spatial and temporal resolutions.

The actual implementation of these models is an important objective, from both theoretical and practical standpoints. The problem is to modify classical algorithms to take into account characteristics extracted from data. For instance, geodesic approaches are mainly based, in their general formulation, on distance and curvature information. This is not necessarily relevant for all kinds of data. One must then determine, from environmental data, semantic characteristics (they can be high level) and modify the model's formulation in such a way that the geodesic propagation is driven by parameters coming from the data. The same is also applicable to motion estimation, which must not rely anymore on gray-level conservation, but on adapted visual indices, together with motion models that tackle fluid motion and turbulence.