LIU et al.: SAR IMAGE SEGMENTATION BASED ON HVS-ANML 4289
Fig. 1. (a) Original synthetic SAR images. (b) Sketch maps. (c) Regional maps, which include aggregated region (in white), structural region (in gray), and
homogeneous region (in black).
2) Using the nonmax suppression and double-threshold-
based connection methods, the curves contained in the
SAR image are extracted.
3) Each extracted curve is represented as a sketch line via an
approximation method.
4) According to the properties of SAR images, a pair of para-
doxical hypotheses is built to evaluate the significance of
each sketch line, and only the significant sketch lines are
preserved to constitute the sketch map of the SAR image.
In the SAR sketch map [see Figs. 1(b) and 2(b)], the black
lines are the sketch lines. Each sketch line consists of several
sketch line segments, which are connected end to end. The
end points of each sketch line segment are marked as the solid
dots. Moreover, the sketch lines in the aggregated scenes are
aggregated, while the sketch lines indicating the edges and line
objects are not aggregated. It means that the sketch lines in dif-
ferent regions usually have different characteristics. According
to these different characteristics, the aggregated degree is used
to classify the sketch lines into aggregated and nonaggregated
sketch lines. Then, the region extractor is implemented on the
aggregated sketch lines to obtain the aggregated region [32]
(the details are introduced in the Appendix). The GSW is
operated on the nonaggregated sketch lines to obtain the struc-
tural region, where GSW is an oriented rectangular window.
Specifically, two sides of GSW are parallel to the sketch line,
and the other two sides are vertical to the sketch line. The re-
maining part of the sketch map is the homogeneous region. The
aggregated, structural, and homogeneous regions constitute the
regional map. According to the obtained regional map, a SAR
image is reasonably divided into aggregated, structural, and
homogeneous regions. It is a kind of sparse representation [33]
of SAR images. Specifically, the aggregated scenes are usually
contained in the aggregated portion. The details are always con-
tained in the structural portion. The regions with a slight change
of gray value are mostly contained in the homogeneous portion.
The sketch maps and regional maps of two synthetic SAR
images and seven real SAR images are shown in Figs. 1 and 2,
respectively. In the sketch maps [shown in Figs. 1(b) and 2(b)],
the black lines are the sketch lines. We can see that the sketch
lines in the real SAR images are more complicated than that
in the synthetic SAR images. The regional maps [shown in
Figs. 1(c) and 2(c)] include three portions. The white, gray, and
black portions are the aggregated, structural, and homogeneous
regions, respectively.
From the above, we can see that the sketch map is a primal se-
mantic space which represents the structure of the SAR image.
The primitive of the sketch map is a sketch line. The regional
map is a higher semantic space, whose primitive is a region.
Moreover, the semantic is based on the framework of Marr’s
vision. Therefore, we name it as hierarchical visual semantic,
which is used to guide the SAR image segmentation. Specifi-
cally, the regional map is used to divide SAR images into aggre-
gated, structural, and homogeneous regions. It can improve the
consistencies of the aggregated land covers in the segmentation
process. The sketch map is used to design GSW and visual se-
mantic rules. It can improve the location accuracy of the edges
and line objects in the segmentation process. Since different
levels of semantics are used to deal with regions with different
characteristics, the consistencies of the regions and the details’
preservations will be achieved simultaneously in the segmen-
tation process. The hierarchical visual semantic and its corre-
sponding semantic representation of SAR images are shown in
Fig. 3. Fig. 3(a) shows the hierarchical visual semantic. The
first level is the original SAR image, the second level is the
sketch map, and the third level is the regional map. In Fig. 3(b),
the regional map is mapped into the SAR image, which is
divided into aggregated, structural, and homogenous regions.