Polar-fisherspace: a linear transformation invariant model for appearance based object recognition
B. H. Shekar, Panduranga Naidu Nagabhushan, D. S. Guru · 2006
In this paper, we present a polar-fisherspace representation scheme for appearance based object recognition. Given an object or a face image, we first transform to its polar coordinate (r, thetas) representation called polar image using the center of mass of the image as the origin. Fourier spectrum of a polar-image is obtained for achieving invariance to linear transformations (in-plane rotation and scale). The transformed polar image is then subjected to the (2D)2-FLD (Nagabhushan et al., 2006) to construct polar-fisherspace. Experiments conducted on COIL-20 dataset and AT&T face dataset show that the proposed polar-fisherspace representation is invariant to linear transformations and hence is suitable for real time practical recognition applications.