Image Processing Reference
In-Depth Information
selected imagery types giving some general notion about correlation between image data
type, rough entropy measure, fuzzy and rough parameters and other distinct segmentation
quality measures.
Application of the presented research material relates to better insight into rough entropy
notion that should improve understanding of this new concept in the context of data analysis
and give way to practical incorporation into other data analysis frameworks. Additionally,
experimental results describe correlation between selected image quality assessment indices.
This type of information is valuable during construction and implementation of the image
segmentation and analysis systems.
Section 11.2 presents general information concerning segmentation and clustering algo-
rithms. In Section 11.4 methods of image segmentation has been presented together with
description of selected validation measures described in Section 11.3 that are used in evalua-
tion experiments. Experimental setup has been presented in Section 11.5 and experimental
results in Section 11.6. Finally, conclusions are drawn with summary of the obtained results.
11.2
Clustering Based Image Segmentation
Segmentation operation is an essential and extremely important preprocessing step in the
majority of image analysis based routines such as computer vision with practical appli-
cations ranging from object extraction and detection, change detection, monitoring and
identification tasks. After the preprocessing stage of image handling routines, with for
example noise removal, smoothing, and sharpening of image contrast, follows the image
segmentation step, and subsequently more specific, high-level analysis is performed such
as depicting objects and regions, and final interpretation of the image or scene. In almost
all areas, the quality of the segmentation step determines the quality of the final image
analysis output. Segmentation process is defined as an operation of image partitioning into
some non-overlapped regions such that each region exhibits homogeneous properties and no
two adjacent regions are homogeneous by means of intensity, color, texture or other rele-
vant features. Additionally, most often other conditions are simultaneously imposed in the
form, for example regions interiors that should be simple and without many small holes,
and each segment boundaries should be comparatively simple and should have spatially
accurate structure.
Image segmentation routines are divided into: histogram-based routines, edge-based rou-
tines, region-merge routines, clustering routines and some combination of the above rou-
tines. Exhaustive overview of the segmentation methods is available in (Fu and Mui, 1981)
and image segmentation evaluation methods in (Zhang, 1996). Edge detection approaches
to image segmentation deal with discovering image locations where sharp changes in grey
level or color are detected. The main di
culty in this type of algorithms is maintenance
of the continuity of detected edges. Segments have always to be enclosed by continuous
edges. However, usually disconnected or isolated edges within areas with more details have
to be combined by using specialized heuristics. Region growing or merging presents an ap-
proach for image segmentation where large continuous regions or segments are detected first
and afterwards, small regions are subjected to merging operation by use of a homogeneity
criteria. Region growing and merging routines are sequential in nature, dependent upon
the order in which regions grow or merge.
Furthermore, algorithms based on combining
Search WWH ::
Custom Search