Project title: Brushlet-Based Adaptive Probabilistic Modeling of Textures

Abstract

This work investigates a probabilistic model of textures using brushlet, i.e., steerable wavelet packet. Compare to wavelet packet, brushlet provides a most concise and precise representation of a texture with all possible directions, frequencies, and locations. A generalized quartic model effectively models the statistics of subband coefficients of brushlets comprising Gaussian, Generalized Gaussian and Bimodal statistics. The optimal partition tree of brushlet for each texture is trained and the model parameters of each subband are estimated using MAP. The model parameters of the optimal partition tree are used as feature vectors for texture classification. The experiments demonstrate our proposed probabilistic modeling of brushlet has excellence performance in analyzing and describing textures for application of classification. Brushlet-based adaptive probabilistic model of texture provides also a feasible solution for rotation invariant texture classification and segmentation.

Preliminary results

Interpretations:
Row 1: Texture patch from image hexholes-1.5.2 (128x128 pixels),
Row 2: Its optimal brushlet decomposition and the optimal partition tree - different colors correspond to the different models automatically selected within each subband: black, gray and white indicate the Gaussian, Generalized Gaussian and Bimodal model, respectively,
Row 3: Histograms from a black, a gray and a white colored subband, respectively.

Last modified: Fri Sep 01 15:02:52 CEST 2006