Gaussian Markov Random Fields ( GMRF ) have been successfully used to model textures. However, they do not provide the best results for classifying self-similar textures. In this paper, we model self-similar textures using a wavelet representation. We show that the detail signal of a Fractional Brownian Motion ( FBM ) is zero mean stationary. The detail signal of the self-similar texture is modeled as a Gaussian Markov Random Field. Texture classification is performed using the parameters of the Gaussian Markov Random Field of the detail signal.