Manuscript [PDF coming soon]

Supplemental Material [PDF]

Abstract

Natural stochastic textures coexist in images with complementary edge type structural elements that constitute the cartoon-type skeleton of an image. Separating texture from the structure of natural image is an important inverse problem in image analysis. The concept behind this decomposition is to extract the textural layer, which conveys fine details and small-scale variations, and separate them from the image macrostructures (edges and contours). In this paper, we propose a variational texture-structure separation scheme. Our approach involves texture modeling by a stochastic field; The 2D fractional Brownian motion (fBm), a non-stationary Gaussian self-similar process, has been shown to be a suitable model for pure natural stochastic textures. We therefore use it as a reconstruction prior to extract the corresponding textural element and show that this separation is crucial for improving the execution of various image processing and computer vision tasks (e.g. clustering, classification and enhancement). Here we show the applicability of this decomposition in image denoising. Lastly, we highlight how manifold-based representation of texture-structure data, can be implemented in extraction of geometric features and thereby construction of a classification space.

Decomposition Results 

Notations