Abstract (eng)
In different imaging scenarios, such as medical and biological applications, the alignment of two or more images of similar objects is of crucial importance. Two such problems are treated in this work, from the point of view of continuous variational models which are then discretized for numerical computations. A common feature of the models presented is the use of nonconvex regularization, in addition to the natural nonconvexity of registration problems.
The first is surface matching, in which the data is given as two different surfaces. In this framework, we consider surfaces embedded in some computational domain and represented by their signed distance functions. Our approach is to consider shell energies penalizing expansion, compression and bending of the first surface, which are simplified using the level set scenario and the geometry of the second surface. For this problem, two models are proposed. The first is a direct approach which effectively encodes the geometry of the situation, while the second formulation is further refined to allow proving weak lower semicontinuity and existence of minimizers, along with efficient numerical computations on adaptive grids.
The second is the estimation of optical flow along a full sequence of images. For it, a novel time regularization along the trajectories of the flow is proposed. It penalizes the convective acceleration of the resulting vector field, instead of the naive time derivative of the Eulerian velocity field. The resulting problem can then be approximated in a semi-implicit fashion by a sequence of linear ones. Numerical results show a marked improvement with respect to just using the time derivative.