Nonlinear Functionals in the Construction of Multiscale Affine Invariants

In this paper we introduce affine invariants based on a multiscale framework combined with nonlinear comparison operations. The resulting descriptors are histograms, which are computed from a set of comparison results using binary coding. The new constructions are analogous to other multiscale affine invariants, but the use of highly nonlinear operations yields clear advantages in discriminability. This is also demonstrated by the experiments, where comparable recognition rates are achieved with only a fraction of the computational load. The new methods are straightforward to implement and fast to evaluate from given image patches.