Texture Segmentation Using Recursive Markov Random Field Parameter Estimation

Abstract: 
An efficient and robust type of unsupervised colour texture segmentation method.

 

The algorithm starts with spectral factorization of an input colour texture mosaic

Natural texture mosaic (marble, sand, grass, stone).

using the Karhunen-Loeve expansion.

Single Karhunen-Loeve factors.

 

Monospectral factors of single texture patches are assumed to be modelled using a Gaussian Markov random field model. The texture segmentation is done in the Markov model parameter space evaluated for each pixel centered image window and for each spectral band. A novel recursive maximum pseudo-likelihood estimation procedure of a Gaussian Markov random field model parameters is derived for the efficient evaluation of segmentation statistics. It can be easily shown that for all possible ranges of contextual support set for the model as well as estimation window sizes the recursive estimation procedure is always faster than the classical approach. For example, for a window of the 50 pixels width and the model with the second order hierarchical contextual support set the non recursive estimate needs one million extra operations.

 

     Optimal segmentation.

 

The method estimates texture model parameters independently in each spectral band and combines them into the multispectral segmentation feature space. Segmentation is done for each feature and single thematic maps are combined together.

Resulting segementation before postprocessing.

 

Reference: 
Haindl, M., "Texture segmentation using recursive Markov random field parameter estimation", Proceedings of the 11th Scandinavian Conference on Image Analysis, Lyngby, Pattern Recognition Society of Denmark, pp. 771-776, June, 1999.