Luminance-Chrominance linear prediction models for color textures: An application to satellite image segmentation

This thesis details the conception, development and analysis of a novel color texture descriptor based on the luminance-chrominance complex linear prediction models for perceptual color spaces. In this approach, two dimensional complex multichannel versions of both causal and non-causal models are developed and used to perform the simultaneous parametric power spectrum estimation of the luminance and the combined chrominance channels of the proposed two channel complex color image. The accuracy and precision of these spectral estimates along with the spectral distance measures ensure the robustness and pertinence of the approach for color texture classification. A luminance-chrominance spectral interference based quantitative measure for the color space comparison is also introduced. The experimental results for different test data sets, in IHLS and L*a*b* color spaces are presented and discussed. These results have shown that the chrominance structure information of the color textured images could get better characterized in L*a*b* color space and hence could provide the better color texture classification results. A Bayesian framework based on the multichannel linear prediction error is also developed for the segmentation of textured color images. The main contribution of this segmentation methodology resides in the robust parametric approximations proposed for the multichannel linear prediction error distribution. These comprised of a unimodal approximation based on the Wishart distribution and a multimodal approximation based on the multivariate Gaussian mixture models. Another novelty of this approach is the fusion of a region size energy term with the conventional Potts model energy to develop a Gibbs random field model of the class label field. This improved label field model is used for the spatial regularization of the initial class label estimates computed through the proposed parametric priors. Experimental results for the segmentation of synthetic color textures as well as high resolution QuickBird and IKONOS satellite images validate the application of this approach for highly textured images. Advantages of using these priors instead of classical Gaussian approximation and improved label field model are evident from these results. They also verify that the L*a*b* color space exhibits better performance among the used color spaces, indicating its significance for the characterization of complex textures through this approach.