Recognition of rotated and scaled textures using 2-D AR modeling and the Fourier-Mellin transform

In this communication, we address the problem of the recognition and classification of rotated and scaled stochastic textures. We propose an extension of previous works, which are mostly based upon the use of the Fourier transform for the derivation of translation invariant statistics. More precisely, we suggest the sequential use of a 2-D high resolution spectral estimate, called Harmonic Mean power spectrum density (HM PSD), which presents a good tradeoff between reliability and complexity, with the Fourier-Mellin transform from which rotational and scaling invariants can be derived. Experimental results on rotated and scaled textures are presented that show the efficiency of this new technique.