Image segmentation using wavelet coefficients and geodesic distance between elliptical distributions for applications in street view

Abstract

The geodesic distance on the manifold of multivariate zero-mean Generalized Gaussian Distributions (GGD) has been shown a strong similarity measure for texture classification. Recent works demonstrates that the GGD can be employed for texture identification in the wavelet domain with more accuracy than other measures, like the Kullback Leibler Divergence. The wavelet coefficients of an image can be grouped considering color and spatial dependence. The Laplacian distribution is one of various possible elliptical distributions and is the choice of this work for modeling these coefficients. A street view application of this technique is presented. First, a wavelet decomposition of the image is done. Then, the coefficients of smaller regions (windows) are grouped, and a Laplacian distribution is computed for each coefficients group at each subband. The geodesic distance between these distributions can be computed. This can be viewed as a similarity measure between the regions of the image, and a spectral clustering is employed, using the k-means method for the segmentation. Thus, regions with different textures, as the streets, can be discriminated from each other. The main contribution of this paper is the use of the geodesic distance between GGDs in a segmentation context.