Sparse set of Features for Texture Discrimination

This package implements the Features mentioned in the PhD thesis of Thomas Brox.

The feature vector is a 5(gray scale) or 15(colored) dimensional vector reflecting the contrast, texture strength and orientation and texture scale for each pixel. Texture scale is represented by the average speed of change of pixel intensity in a Total Variation framework while the texture strength and orientation are computed from 3 distinct components of the structure tensor undergone a nonlinear coupled isotropic matrix valued diffusion. The feature vector can directly be used in a texture segmentation framework.
You need to download the Nonlinear Coupled Diffusion package (submission 27604 available at http://www.mathworks.com/matlabcentral/fileexchange/27604-nonlinear-coupled-diffusion) to run this code.
The code is commented and the definitions of the input/output variables and usages are mentioned in the header of the discriminative_texture_feature.m. A sample script test_discriminative_texture shows the usage of the code as well as the usefulness of the gaussian regularization for speedups in the nonlinear diffusion process.
Note that you need to compile the mex file: thomas_mex.cpp before the first usage.