analysis (OBIA) approach, which generally includes low-, medium-, and high-level
image processing subtasks:
• Preprocessing steps (geometric and radiometric corrections, filtering, scalespace image simplification, edge detection, band math expression computations, etc.)
• Image segmentation (in order to produce single-level or multilevel hierarchies
of primitive objects within the image space)
• Computation of image object properties based on spectral, shape, topological,
and context features
• Definition of object-oriented class hierarchy and representation of knowledge
through rules or through training with samples
• Classification (learning techniques or rule-based systems to perform the classification task)
• Accuracy assessment in order to derive the quality of the resulting classification
• Vectorization steps (create the output to spatial databases and integrate the
information to thematic maps).
The developed approach is integrating certain advanced computer vision and
machine learning methods for implementing the above tasks. Although the proposed
methodology is following the above general OBIA approach, which is also available
in several proprietary and free software solutions, the originality of the developed
framework lies in the novel approach that excludes the parameter tuning step, in the
robustness of the multiscale hybrid (edge-based and region-based) segmentation
algorithm, in the extensive use of nonlinear scale-space representations for image
simplification purposes, and in the integration with kernel-based machine learning
methods for classification.
Briefly the developed framework consists of the following steps: First, for every
band of the initial image, a scale-space representation is generated using the anisotropic
morphological leveling (AML) formulation (Karantzalos et al., 2007). The supported
type of the imagery can be up to double precision and of any number of bands. A feature
extraction step is then applied on the scale-space stack. For this step two algorithms were
tested and are presented here: the Canny edge detector (Canny, 1986) and the line
segment detector (LSD) (Von Gioi et al., 2010). A multiscale segmentation algorithm is
applied afterward which is able to integrate the simplified scale-space stack along with
the corresponding edge information. During this step primitive image objects are
formed. The procedure starts from single pixel, and through pairwise merges bounded
by edge information, several levels of image objects are produced. The multiscale object
hierarchies are been constructed without any parameter tuning. In a similar way with
Tzotsos et al. (2011), here the edge features are produced without tuning the edge
extraction parameters (Figure 9.1). Last in the processing order comes a dual classification procedure using a support vector machine classifier. The first classification is
performed on all scale-space representations and their corresponding segmentations,
while the second optimal one is performed after an interim accuracy assessment.
176
MULTISCALE SEGMENTATION AND CLASSIFICATION
image processing subtasks:
• Preprocessing steps (geometric and radiometric corrections, filtering, scalespace image simplification, edge detection, band math expression computations, etc.)
• Image segmentation (in order to produce single-level or multilevel hierarchies
of primitive objects within the image space)
• Computation of image object properties based on spectral, shape, topological,
and context features
• Definition of object-oriented class hierarchy and representation of knowledge
through rules or through training with samples
• Classification (learning techniques or rule-based systems to perform the classification task)
• Accuracy assessment in order to derive the quality of the resulting classification
• Vectorization steps (create the output to spatial databases and integrate the
information to thematic maps).
The developed approach is integrating certain advanced computer vision and
machine learning methods for implementing the above tasks. Although the proposed
methodology is following the above general OBIA approach, which is also available
in several proprietary and free software solutions, the originality of the developed
framework lies in the novel approach that excludes the parameter tuning step, in the
robustness of the multiscale hybrid (edge-based and region-based) segmentation
algorithm, in the extensive use of nonlinear scale-space representations for image
simplification purposes, and in the integration with kernel-based machine learning
methods for classification.
Briefly the developed framework consists of the following steps: First, for every
band of the initial image, a scale-space representation is generated using the anisotropic
morphological leveling (AML) formulation (Karantzalos et al., 2007). The supported
type of the imagery can be up to double precision and of any number of bands. A feature
extraction step is then applied on the scale-space stack. For this step two algorithms were
tested and are presented here: the Canny edge detector (Canny, 1986) and the line
segment detector (LSD) (Von Gioi et al., 2010). A multiscale segmentation algorithm is
applied afterward which is able to integrate the simplified scale-space stack along with
the corresponding edge information. During this step primitive image objects are
formed. The procedure starts from single pixel, and through pairwise merges bounded
by edge information, several levels of image objects are produced. The multiscale object
hierarchies are been constructed without any parameter tuning. In a similar way with
Tzotsos et al. (2011), here the edge features are produced without tuning the edge
extraction parameters (Figure 9.1). Last in the processing order comes a dual classification procedure using a support vector machine classifier. The first classification is
performed on all scale-space representations and their corresponding segmentations,
while the second optimal one is performed after an interim accuracy assessment.
176
MULTISCALE SEGMENTATION AND CLASSIFICATION