UDCT complex coefficient statistics based rotation invariant texture characterization

We propose a discriminative texture feature based on a recent discrete implementation of the curvelet transform, namely the uniform discrete curvelet transform (UDCT). Several approaches including either statistical methods or spectral methods have been considered to describe the characteristics of textured surface. Anyhow, most of the proposed texture features are sensitive to rotation variations. In this paper, statistical properties of complex subband coefficients are captured more accurately by using the efficiency of the UDCT in extracting edge and linear information, and an accurate statistical modeling of complex coefficient distributions based on the bivariate generalized Gaussian distribution. Texture classification performances are carried out to investigate the robustness of the proposed descriptor. The results show that the classification rate of the proposed feature outperforms these of compared feature extraction methods considering marginal distributions, while achieving the invariance property to rotated image patterns.