The MSEG algorithm was improved in order to be able to integrate edge
information (as a constraint) during the segmentation procedure. The goal was to
design a more robust and generic segmentation procedure that would be able to
take into account advanced edge and line features. In particular, the region merging
algorithm starts by selecting initialization points throughout the image using the
SPE (start point estimation) module (Tzotsos and Argialas, 2006) and a queue of
pixels is created in order to be able to achieve reproducibility. Then, iterative
pairwise fusions occur within the image space, starting from single-pixel objects,
in a way that local heterogeneity is minimized (for color and shape criteria).
During this pairwise merging of image objects, the edge information is used as a
boundary. Two adjacent pixels will not be merged into an object if one or both
reside on top of an edge. After the first pass of the region merging procedure,
image objects of one or two pixels exist, with edge pixels being constrained and
not merged to each other. During the following passes edge objects (still single
pixels) are not merged, thus not permitting object merging between image regions
that are separated by a line or a continuous edge feature. After several passes
(iterations) converging of the algorithm occurs and no more object merging is
performed due to scale parameter. At this point the edge objects are still intact by
the region merging procedure, thus binding the procedure into respecting edge
features. Finally a last iteration of the algorithm is forced on edge objects only and
a selection is made to which neighboring object they should be merged based on
local heterogeneity. This step is taking advantage of the fact that both Canny and
LSD features are one pixel wide, and thus edge objects are always capable of
merging with nonedge objects.
A certain novelty of the developed segmentation process is that it does not use an
edge penalty model for the edge compensation as presented in other approaches
(Kermad and Chehdi, 2002; Cufi et al., 2003; Yu et al., 2012); rather it uses a
topological constraint effective throughout the region merging procedure. The results
of this enhancement is presented in Figures 9.1e,j,o,t showing very promising
segmentation results. A scale parameter of value 100 was used for all tests on Figure
9.1, showing that scale-space representation effectively provides the scale of the
obtained objects. More results are presented and discussed in the following sections.
9.3.3 Kernel-Based Classification
For the developed approach, an Support Vector Machine (SVM) classification
scheme (Vapnik, 1998; Tzotsos, 2006; Tzotsos and Argialas, 2008; Tzotsos et al.,
2011) was employed. After the multiscale segmentation which is constrained by
edge information, image objects were extracted and object properties were computed forming the feature space of the classification step. For each primitive image
object, spectral, shape, and spatial properties (such as mean band values, standard
deviation, Gray-Level Co-occurrence Matrix (GLCM) texture features, area,
perimeter, compactness, and number of neighbors) were extracted by the topological model used to handle the object topology. This model was proposed by
METHODOLOGY
181
information (as a constraint) during the segmentation procedure. The goal was to
design a more robust and generic segmentation procedure that would be able to
take into account advanced edge and line features. In particular, the region merging
algorithm starts by selecting initialization points throughout the image using the
SPE (start point estimation) module (Tzotsos and Argialas, 2006) and a queue of
pixels is created in order to be able to achieve reproducibility. Then, iterative
pairwise fusions occur within the image space, starting from single-pixel objects,
in a way that local heterogeneity is minimized (for color and shape criteria).
During this pairwise merging of image objects, the edge information is used as a
boundary. Two adjacent pixels will not be merged into an object if one or both
reside on top of an edge. After the first pass of the region merging procedure,
image objects of one or two pixels exist, with edge pixels being constrained and
not merged to each other. During the following passes edge objects (still single
pixels) are not merged, thus not permitting object merging between image regions
that are separated by a line or a continuous edge feature. After several passes
(iterations) converging of the algorithm occurs and no more object merging is
performed due to scale parameter. At this point the edge objects are still intact by
the region merging procedure, thus binding the procedure into respecting edge
features. Finally a last iteration of the algorithm is forced on edge objects only and
a selection is made to which neighboring object they should be merged based on
local heterogeneity. This step is taking advantage of the fact that both Canny and
LSD features are one pixel wide, and thus edge objects are always capable of
merging with nonedge objects.
A certain novelty of the developed segmentation process is that it does not use an
edge penalty model for the edge compensation as presented in other approaches
(Kermad and Chehdi, 2002; Cufi et al., 2003; Yu et al., 2012); rather it uses a
topological constraint effective throughout the region merging procedure. The results
of this enhancement is presented in Figures 9.1e,j,o,t showing very promising
segmentation results. A scale parameter of value 100 was used for all tests on Figure
9.1, showing that scale-space representation effectively provides the scale of the
obtained objects. More results are presented and discussed in the following sections.
9.3.3 Kernel-Based Classification
For the developed approach, an Support Vector Machine (SVM) classification
scheme (Vapnik, 1998; Tzotsos, 2006; Tzotsos and Argialas, 2008; Tzotsos et al.,
2011) was employed. After the multiscale segmentation which is constrained by
edge information, image objects were extracted and object properties were computed forming the feature space of the classification step. For each primitive image
object, spectral, shape, and spatial properties (such as mean band values, standard
deviation, Gray-Level Co-occurrence Matrix (GLCM) texture features, area,
perimeter, compactness, and number of neighbors) were extracted by the topological model used to handle the object topology. This model was proposed by
METHODOLOGY
181