Invariant object recognition with discriminant features based on local fast-Fourier Mellin transform

We describe an approach to scale-, orientation- and shift-invariant object recognition. The modulo of locally computed Fourier-Mellin descriptors serve as features that describe local image-patches scale- and orientation-invariant. Those features can be efficiently computed w.r.t. each image location, thus enabling positional invariance as well. Based on those features, we use a two-step procedure for locating and subsequently identifying previously learnt objects. First, all known objects are searched for in parallel in the principle components domain using a probabilistic similarity measure. Hereafter, possible object locations are further examined using Fisher's discriminant analysis, thus enabling multiobject identification in one step. A spin-off from the principle component analysis enables representation-based feature selection, which in turn reduces the computational burden of feature generation.