Biomedical Engineering Reference
In-Depth Information
were used in [76] for better orientation adaptivity. To further extend the flexibil-
ity of the spatial-frequency analysis, a multiwavelet packet, combining multiple
wavelet basis functions at different expansion levels, was used in [77]. An
M
-
band wavelet expansion, which differs from a dyadic wavelet transform in the
fact that each expansion level contains
M
channels of analysis, was used in [78]
to improve orientation selectivity.
Quality and accuracy of segmentation ultimately depend on the selection of
the characterizing features. A simple feature selection can use the amplitude
of the wavelet coefficients [76]. Many multiscale texture segmentation methods
construct the feature vector from various local statistics of the wavelet coeffi-
cients, such as its local variance [73, 79], moments [80], or energy signature [74,
78, 81]. Wavelet extrema density, defined as the number of extrema of wavelet
coefficients per unit area, was used in [77]. In [75], a 1D envelope detection was
first applied to the wavelet packets coefficients according to their orientation,
and a feature vector was constructed as the collection of envelope values for
each spatial-frequency component. More sophisticated statistical analyses in-
volving Bayesian analysis and Markov random fields (MRF) were also used to
estimate local and long-range correlations [82, 83]. Other multiscale textural fea-
tures were also reported, for example
χ
2
test and histogram testing were used
in [84], “Roughness” based on fractal dimension measurement was used in [85].
Texture-based segmentation is usually achieved by texture classification.
Classic classifiers, such as the minimum distance classifier [28], are easier to
implement when the dimension of the feature vector is small and the groups
of samples are well segregated. The most popular classification procedures re-
ported in the literature are the K-mean classifier [73, 75, 76, 78, 79, 81, 85] and
the neural networks classifiers [27, 74, 80, 82].
As an example, we illustrate in Fig. 6.23 a texture-based segmentation method
on a synthetic texture image and a medical image from a brain MRI data set.
The algorithm used for this example from [75] uses the combination of
wavelet
packets frame
with Lemari e-Battle filters,
multiscale envelope
features, and a
K-mean
classifier.
6.4.2 Wavelet Edge Detection and Segmentation
Edge detection plays an important role in image segmentation. In many cases,
boundary delineation is the ultimate goal for an image segmentation and a good
