Next: Unsupervised segmentation of textures
Up: Indexing and segmentation
Previous: Estimation of the model
In this work we have chosen the Kullback distance
between two Gaussian models of textures :
Clearly, one is interested in being able to define a distance measure,
which is invariant to translation and rotation.To achieve the desired invariance, we have implemented the following
procedure:
- Estimate the parametric model of each texture patch.
- For textures with harmonic components, the texture is
first rotated so that its dominant harmonic component is aligned with
a predefined direction (the x-axis, for example).
- For textures with evanescent components, the dominant evanescent
component is considered the one whose modulating 1-D
purely-indeterministic processes have maximal variance. Thus, the texture is first rotated so that its dominant evanescent component has
.
- For textures with harmonic components where the phase of the
dominant component is not zero, we crop a sub-picture of the original in which the phase of the dominant component is zero.
- Re-estimate the texture parameters to find the parameters of
all the model components.
Figure:
Textures ordered using the Kullback distance. The test texture is structured, dominantly harmonic.
|
Test Texture |
|
The Kullback distance is computationally expensive. Therefore we
derived a hierarchical approach :
- if both textures are deterministic : the
distance becomes the euclidean norm between spectral
densities
- if both textures are purely random : the distance is
reduced to the Itakura-Saito distance
- if one texture is purely random and one texture is
deterministic : the distance is infinity
Figure:
Textures ordered using the hierarchical distance. The test texture is structured, dominantly harmonic.
|
Test Texture |
|
Next: Unsupervised segmentation of textures
Up: Indexing and segmentation of
Previous: Estimation of the model
Radu Stoica
1999-05-21