Spectral control of viscous alignment for deformation invariant image matching

Abstract

We present a new approach to deformation invariant image matching. Our approach retains the broad range of linear and nonlinear deformations that viscous alignment methods can model, but introduces a selectivity that is necessary for recognition. Our method models viscous kernels with an over-complete filter basis. The basis is parameterized with a single scalar parameter, the spectral radius r, which selects deformations ranging in complexity from tranlations to "turbulence." The spectral radius is used for cascaded alignment starting from low deformation frequencies and finishing with high deformation frequencies. Cascaded alignment makes deformation invariant matching for recognition feasible and efficient. Because spectral radii map directly to deformation complexity, their contributions are selectively weighed to calculate the template-target similarity. In this way, our model can distinguish deformations by their relevance to recognition, without losing the flexibility of viscous alignment for handling nonlinear deformations. Our approach is applied to recognize flexible bodies of animals, and results indicate that the method is very promising.