Image Processing Reference
In-Depth Information
edge-based and region-based techniques are capable to exploit the complementary nature
of edge and region information. Additionally, many segmentation techniques make use of
particular data analysis approaches such as neural networks, fuzzy computing, evolutionary
computing, multiscale resolution techniques and morphological analysis. Some segmenta-
tion methods are based on unique frameworks, such as active contour models, active shape
models and watersheds. The active shape model (ASM) presents a particular structure for
finding the object boundary in images. In this framework, different image features and
search strategies are applied that subsequently create a vast range of ASM algorithms. Into
this framework-based segmentation approaches falls clustering with rough entropy based
segmentation quality measure that is the subject of this chapter.
11.2.2 Image Clustering
Image clustering algorithms are widespread high quality procedures applied in many areas
of image processing and image analysis. High robustness of general data clustering schemes
have been successfully incorporated in many image segmentation routines. Data clustering
depends on partitioning data objects into disjoint groups that internally consist of simi-
lar objects and externally objects from different groups are mutually dissimilar as much
as possible in the given data set. Data clustering algorithms are most often divided into
hierarchical and partitioning routines. In the case of hierarchical routines, clustered data
are arranged into groups that create a tree-like hierarchy. Objects from the given group are
similar to each other relative to a predefined similarity measure as opposed to objects from
different groups. However, most often data groups or clusters are the part of hierarchical
level structure that can be considered as data grouping on different granularity or similar-
ity level. On the other hand, data partitioning clustering most often depends on partition
of data into predefined number of clusters on the basis of optimization of some objective
function that reflects internal data similarity. In the group of data partitioning clustering
methods there are many methods that are capable of creating high quality clustering solu-
tions. In this group most prominent are k -means clustering schemes together with numerous
modifications and different data analysis frameworks such as fuzzy or rough k -means algo-
rithms. The necessity for improving basic clustering functionality in, for example, k -means
algorithms, stems from the fact that this type of algorithms most often determines solutions
that are local optima of the objective function. Rough entropy framework proposed recently
in (Malyszko and Stepaniuk, 2008) has been aimed at incorporation of uncertainty notion
in the clustering process and taking advantage of rough set data analysis methods in the
clustering setting.
11.3
Image Segmentation Evaluation
11.3.1 Segmentation Quality Measures
In supervised approaches to segmentation and classification routines, definition of segmen-
tation quality by measures such as accuracy and precision is quite straightforward by means
of ground-truth ideal images. On the other hand, during clustering routines, this kind of
indispensable segmentation quality assessment not always is feasible because ground-truth
image is not present or is di cult to obtain. In the present study, the following segmenta-
tion quality measures were taken into account as cluster validity indices. Data clustering
as data grouping routine (together with other grouping algorithms such as for example
data thresholding) presents unsupervised process that finally requires some sort of quality
evaluation of generated data groups or clusters. This requirement can be satisfied by using
 
 
Search WWH ::




Custom Search