et al., 2012; Ouzounis et al., 2012), studies on a varying stopping time (Gilboa, 2008),
and studies on the behavior on corner and other local descriptors (Zhong et al., 2009;
Jiang et al., 2011; Kimmel et al., 2011; Xu et al., 2012).
9.2.3 Edge-Based Segmentation
Extracting primitives (i.e., contours, edges, lines, etc.) is a basic low-level operation
in the human visual system. Along with the aim to understand and simulate the
human vision, the importance of building up computational models for the
perception of primitives is a major component in many applications of computer
vision, such as object/pattern recognition, robot vision, remote sensing, and
medical image analysis.
These primitives give important information about the geometric content of
images. Most of the Earth surface objects and in particular most man-made objects
are made of flat surfaces with certain geometric features. In addition, many shapes can
be described roughly or in detail with edge and line primitives. Therefore, edge or line
segments can be used as a low-level feature description in order to extract information
from images and can serve as the basic tool to analyze and detect more complex
shapes (Von Gioi et al., 2010; Papari and Petkov, 2011; Wang and Oliensis, 2010;
Chia et al., 2012). In the context of scale-space representations, image primitives may
support more stable and efficient representations since their description can be
independent of the object size. Edge and line primitives, defined mainly by the
object geometric properties, allow the robust and efficient feature comparison in
various scales. The latter is of major importance due to large appearance variations of
object instances belonging to the same class.
However, even the recent more sophisticated edge and line detectors cannot
produce connected segments and suffer from the Earth surface complexity pictured in
images, shadows, occlusions, and so on. Therefore, recent efforts are trying to merge
the advantages of edge/line detection and image segmentation techniques in order to
produce connected object contours/boundaries and a comprehensive object description (Pavlidis and Liow, 1990; Kermad and Chehdi, 2002; Cufi et al., 2003). Certain
primitive combinations have been proposed in order to describe more efficiently
object boundaries (Chia et al., 2012; Klonus et al., 2012). Another recent study
proposed a region-based unsupervised segmentation and classification algorithm
which included the computation of an edge strength model (Yu et al., 2012). This
edge penalty model improved segmentation performance by preserving segment
boundaries.
9.3 METHODOLOGY
The main objective here was to design the overall framework in order to be generic,
robust, and able to process effectively a wide variety of remote sensing data, such as
hyperspectral and multispectral data from ground, aerial, and space borne sensors,
radar data, and digital elevation models. It is based upon the object-based image
METHODOLOGY
175
and studies on the behavior on corner and other local descriptors (Zhong et al., 2009;
Jiang et al., 2011; Kimmel et al., 2011; Xu et al., 2012).
9.2.3 Edge-Based Segmentation
Extracting primitives (i.e., contours, edges, lines, etc.) is a basic low-level operation
in the human visual system. Along with the aim to understand and simulate the
human vision, the importance of building up computational models for the
perception of primitives is a major component in many applications of computer
vision, such as object/pattern recognition, robot vision, remote sensing, and
medical image analysis.
These primitives give important information about the geometric content of
images. Most of the Earth surface objects and in particular most man-made objects
are made of flat surfaces with certain geometric features. In addition, many shapes can
be described roughly or in detail with edge and line primitives. Therefore, edge or line
segments can be used as a low-level feature description in order to extract information
from images and can serve as the basic tool to analyze and detect more complex
shapes (Von Gioi et al., 2010; Papari and Petkov, 2011; Wang and Oliensis, 2010;
Chia et al., 2012). In the context of scale-space representations, image primitives may
support more stable and efficient representations since their description can be
independent of the object size. Edge and line primitives, defined mainly by the
object geometric properties, allow the robust and efficient feature comparison in
various scales. The latter is of major importance due to large appearance variations of
object instances belonging to the same class.
However, even the recent more sophisticated edge and line detectors cannot
produce connected segments and suffer from the Earth surface complexity pictured in
images, shadows, occlusions, and so on. Therefore, recent efforts are trying to merge
the advantages of edge/line detection and image segmentation techniques in order to
produce connected object contours/boundaries and a comprehensive object description (Pavlidis and Liow, 1990; Kermad and Chehdi, 2002; Cufi et al., 2003). Certain
primitive combinations have been proposed in order to describe more efficiently
object boundaries (Chia et al., 2012; Klonus et al., 2012). Another recent study
proposed a region-based unsupervised segmentation and classification algorithm
which included the computation of an edge strength model (Yu et al., 2012). This
edge penalty model improved segmentation performance by preserving segment
boundaries.
9.3 METHODOLOGY
The main objective here was to design the overall framework in order to be generic,
robust, and able to process effectively a wide variety of remote sensing data, such as
hyperspectral and multispectral data from ground, aerial, and space borne sensors,
radar data, and digital elevation models. It is based upon the object-based image
METHODOLOGY
175