9.3.1 Scale-Space Filtering
The first step in the developed approach is the construction of the nonlinear scalespace representation in order to elegantly simplify raw data. Anisotropic diffusion
methods are used widely in computer vision applications to simulate the filtering
procedures that are performed in the human vision system. Such methods provide
robust simplification of images without the loss of important information such as
edges that are of high importance for higher level processing algorithms. Especially in
OBIA applications, where very high resolution data are usually processed, it is very
important to simplify the complexity of the initial data and provide a multiscale
representation since different features of the image reside in different scales.
For this preproccesing step, the AMLs (Karantzalos et al., 2007) were incorporated
in the processing scheme. Anisotropic morphological levelings are a combination of
morphological levelings with anisotropic markers and are employed in order to
achieve better segmentation results, reduce the heterogeneity of image data, reduce
oversegmentation, and create accurate image objects. Figure 9.1 shows that by
creating a series of simplified data scale-space filtering leads to a multiscale
segmentation (without tuning any segmentation parameters such as texture, color,
shape, etc.).
Starting from the initial image and for every available band, a scale-space
representation was generated using the AML formulation. Using iterative anisotropic
morphological operations and increasing scales (10, 50, 100, 500, 1000) a scale-space
cube (3D) representation was constructed from each initial band. The result of this
step was a scale-space stack with simplified versions of the raw data. Note that during
this process edge information was preserved in all scales, contrary to isotropic (e.g.,
Gaussian) filtering that loses edge information as scale increases (Figures 9.1f,g,k,l,p,
q). In the following sections further analysis of Figure 9.1 will follow.
9.3.2 Multiscale Segmentation Based on Advanced Edge Features
Edge and line features were computed for every image in the scale-space stack. Edge
information was obtained from the standard Canny detector (Canny, 1986) and the
recent LSD (Von Gioi et al., 2010, 2012).
The Canny edge detector was employed in order to provide primitive edge features
that were integrated to the implemented region merging algorithm. Throughout
this research, the variance parameter of the Canny detector was set stable to 5. In
Figures 9.1e,n,s results from the application of the Canny edge detection on scalespace images is shown. The LSD is a linear-time detector giving subpixel accurate
results. The LSD algorithm starts by computing the level line angle at each pixel to
produce a level line field, that is, a unit vector field such that all vectors are tangent
to the level line going through their base point. Then, this field is segmented into
connected regions of pixels that share the same level line angle up to a certain
tolerance. These connected regions are called line support regions (Von Gioi et al.,
2012). Each line support region (a set of pixels) is a candidate for a line segment. The
principal inertial axis of the line support region is used as the main rectangle direction.
METHODOLOGY
177
The first step in the developed approach is the construction of the nonlinear scalespace representation in order to elegantly simplify raw data. Anisotropic diffusion
methods are used widely in computer vision applications to simulate the filtering
procedures that are performed in the human vision system. Such methods provide
robust simplification of images without the loss of important information such as
edges that are of high importance for higher level processing algorithms. Especially in
OBIA applications, where very high resolution data are usually processed, it is very
important to simplify the complexity of the initial data and provide a multiscale
representation since different features of the image reside in different scales.
For this preproccesing step, the AMLs (Karantzalos et al., 2007) were incorporated
in the processing scheme. Anisotropic morphological levelings are a combination of
morphological levelings with anisotropic markers and are employed in order to
achieve better segmentation results, reduce the heterogeneity of image data, reduce
oversegmentation, and create accurate image objects. Figure 9.1 shows that by
creating a series of simplified data scale-space filtering leads to a multiscale
segmentation (without tuning any segmentation parameters such as texture, color,
shape, etc.).
Starting from the initial image and for every available band, a scale-space
representation was generated using the AML formulation. Using iterative anisotropic
morphological operations and increasing scales (10, 50, 100, 500, 1000) a scale-space
cube (3D) representation was constructed from each initial band. The result of this
step was a scale-space stack with simplified versions of the raw data. Note that during
this process edge information was preserved in all scales, contrary to isotropic (e.g.,
Gaussian) filtering that loses edge information as scale increases (Figures 9.1f,g,k,l,p,
q). In the following sections further analysis of Figure 9.1 will follow.
9.3.2 Multiscale Segmentation Based on Advanced Edge Features
Edge and line features were computed for every image in the scale-space stack. Edge
information was obtained from the standard Canny detector (Canny, 1986) and the
recent LSD (Von Gioi et al., 2010, 2012).
The Canny edge detector was employed in order to provide primitive edge features
that were integrated to the implemented region merging algorithm. Throughout
this research, the variance parameter of the Canny detector was set stable to 5. In
Figures 9.1e,n,s results from the application of the Canny edge detection on scalespace images is shown. The LSD is a linear-time detector giving subpixel accurate
results. The LSD algorithm starts by computing the level line angle at each pixel to
produce a level line field, that is, a unit vector field such that all vectors are tangent
to the level line going through their base point. Then, this field is segmented into
connected regions of pixels that share the same level line angle up to a certain
tolerance. These connected regions are called line support regions (Von Gioi et al.,
2012). Each line support region (a set of pixels) is a candidate for a line segment. The
principal inertial axis of the line support region is used as the main rectangle direction.
METHODOLOGY
177