Image Processing Reference
In-Depth Information
16
Grouping and Unsupervised Region Segregation
In this chapter we present the elements of unsupervised texture boundary estimation
to address automatic grouping of regions in images. The grouping is based on the
assumption that a set of feature vectors is densely available for the image. These
can be viewed as layers of images, each representing a property of the image. A
simple example is an HSV-type color image, where the first layer represents the hue,
the second layer saturation, and the third layer the brightness. However, the feature
images need not come from imaging sensors. They could be local image properties,
densely computed from ordinary gray value images, such as the texture properties:
mean, variance, and direction. A set of N dense feature layers is equivalent to a 2D
grid on which the image takes N -dimensional vector values.
The presentation in this chapter focuses on group formation and boundary esti-
mation in 2D images having tensors as pixel values, but it does not presuppose or
prescribe a specific set of features, neither imaged nor computed, as this is applica-
tion dependent. The value of a pixel is then represented by an array, each element of
which over the entire image grid can be considered as a separate image, also called
a feature image , with real pixel values. First an issue critical to the grouping process
of images is discussed, the uncertainty principle . The presence of noise in the fea-
tures causes a class overlap that can be reduced in a multiresolution pyramid. The
uncertainties in boundaries can be reduced by means of butterfly-shaped smoothing
filters, adaptive to boundary directions. We give further precision to these matters
in the subsequent sections, where the discussions follow the same principles as the
studies in [195], [207], and [227].
16.1 The Uncertainty Principle and Segmentation
Uncertainty in computed image features is similar to the uncertainty principle in
physics. A wave describing a particle cannot be highly concentrated simultaneously
in position and frequency. In unsupervised image segmentation, both position (class
boundaries) and prototype features (local spectral properties) have to be determined.
Finding boundaries is dependent on prototypes and vice versa because:
Search WWH ::




Custom Search