Oriented Patterns Synthesis

This paper explores two different methods to capture the local anisotropy of locally parallel textures. The corresponding models allow for both an analysis and a synthesis of oriented patterns. The first method models such textures as local oscillations that can be analyzed by a windowed Fourier transform. The resulting model can be sampled by alternating projections that enforce the consistency of the local phase with the estimated spectrum of the input texture. The second method models the texture over the domain of local structure tensor. These tensors encode the local energy, anisotropy and orientation of the texture. The synthesis is performed by matching the multiscale histograms of the tensor fields. Figure 1: Examples of locally parallel textures. This paper exposes two models for analyzing and synthesizing natural textures that exhibit a local anisotropy. In section 3 we propose a model based on a local Fourier expansion for the analysis and on iterated projections for the synthesis of the phase function. In section 4, we propose a statistical model that captures the variations of the orientation over a tensor domain.