Classification of Rotated and Scaled Textures using HMHV Spectrum Estimation and the Fourier-Mellin Transform

In the framework of image modeling for texture analysis, we propose the combination of the new parametric 2-D spectrum estimation method called HMHV (Harmonic Mean Horizontal Vertical) and the Fourier-Mellin transform. This latter technique allows the calculation of a set of texture descriptors from a 2-D spectrum estimate which is invariant under rotation and scaling. A comparison of the HMHV and the «standard» parametric HM (Harmonic Mean) methods on a synthetic and natural stochastic textures shows that HMHV method presents almost no spurious peaks and is quite isotropic. By performing the classification of a set of 60 images divided in 12 texture classes, descriptors computed with the HMHV method provide better results than those computed with the HM method.