Metamorphosis of images in reproducing kernel Hilbert spaces (Q295375)

The paper deals with the metamorphosis problem, i.e., ``near-diffeomorphic'' methods that allow one to compare non-diffeomorphic objects within a well-specified Riemannian geometric framework. In the paper, the authors generalize previously known results for image metamorphosis and introduce a new shooting method for computing minimizers of the image metamorphosis matching functional in the case where the images have some degree of smoothness. The paper is divided into two parts. In the first part the authors provide a formal presentation of the proposed approach. It starts with the introduction of basic notation and assumptions. Moreover, the metamorphosis variational problem and associated optimality equations are introduced. Then, a description of a family of singular solutions that satisfy the optimality equations is given. These solutions are then reinterpreted as the solutions that arise from a relaxation of the original problem replacing the infinite-dimensional boundary conditions in image space with a finite number of constraints. Next, the authors describe the numerical algorithm for the solution of the relaxed problem. The first part ends with numerical examples. The second part of the paper is devoted to formal proofs of the presented framework. The authors prove the existence of solutions of the boundary value problem, the optimality conditions and the existence of solutions of the initial value problem.