A Computational Approach to Fisher Information Geometry with Applications to Image Analysis

We develop a computational approach to non-parametric Fisher information geometry and algorithms to calculate geodesic paths in this geometry. Geodesics are used to quantify divergence of probability density functions and to develop tools of data analysis in information manifolds. The methodology developed is applied to several image analysis problems using a representation of textures based on the statistics of multiple spectral components. Histograms of filter responses are viewed as elements of a non-parametric statistical manifold, and local texture patterns are compared using information geometry. Appearance-based object recognition experiments, as well as region-based image segmentation experiments are carried out to test both the representation and metric. The proposed representation of textures is also applied to the development of a spectral cartoon model of images.