Localized Radon Polar Harmonic Transform (LRPHT) Based Rotation Invariant Analysis of Textured Images

2. Related Work

The first algebraic moment was introduced by famous German Mathematician P.A Gordon and D. Hilbert. Kashyap et al. [1986] developed a 2-D circular symmetric auto regressive model to get rotation invariance. A circular symmetric matrix was developed for each pixel point on a circle. The drawback of SAR model is that they are only defined on one circle only. This results in inaccurate description between a pixel and its neighbor. Ojala et al. [2000] proposed a rotation invariant texture classification using local binary pattern. Limited subsets of uniform patterns were used instead of all rotation invariant patterns. However, they overlook the global information of texture. A model based texture classification using spatial autocorrelation function was proposed in [Campisi et al., 2004]. Two steps were used for feature set extraction. The first one is the estimation of binary excitation and in the second step extracting of moment basis to characterize the Autocorrelation Function (ACF) of the binary excitation. It shows highly accurate classification in the presence of noise. Davis [1981] uses a polarogram to get invariant texture features. Co-occurrence matrix was computed prior to the polarogram. Estimation of number of polarogram to be used to compute the co-occurrence matrix is an open problem associated with polarogram. Mao & Jain [1992] developed a rotation invariant SAR (RISAR) model. It gives better results compare to SAR based model. The drawback associated with this model is in two aspects. One is the improper neighbor size and the second is necessity of choosing appropriate window size.