In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
♦
548
(g) (h)
Fig. 2. Flowchart of the proposed model, (a) original image; (b) edges probability map (Ratio response map); (c) over
segmentation results (superpixels); (d) graph structure for GraphCut segmentation; (e) GraphCut segmentation; (f) segmentation
results; (g) map of linear target and distance map from pre-pixels to linear targets; (h) linear target prior maps. The iteration
strategy is marked with dotted lines.
consistency among homogenous regions, but can hardly
consider the consistency along the linear targets like roads and
rivers. 2) Superpixels of Pre-segmentation are captured on
edges probability maps instead of original images. Since shapes
of linear targets are always the boundaries of superpixels. In
this case, more information of edges can be used for
classification process. 3) Iterative MRF description model is
more likely to remove noise in classification map compared
with standard MRF model.
2. RELATED WORK
2.1 Edge detection
Ratio line detector D1 (F.Tupin, 1996) is derived from a
coupling of two ratio edge detectors on both sides of a region
(as shown in fig.3.a). Due to multiple responses to a structure,
detector D1 is not accurate enough to locate the edges. Cross
correlation line detector D2 (F.Tupin, 1996) utilizes variances
of regions to improves locating accuracy but with higher
missing alarm ratio. Tupin (F.Tupin, 1996) merged the
information from both D1 and D2 in 8 orientations.
Fig.3. (a) Template and 8 orientations of template used in
ratio edge detection, (b) Results of ratio edge detection, (c)
Grouping results, (d) Linear targets detection results
2.2 AdaBoost based MRF Model
MRF is a type of classical discriminative model. Given an
image l = {¿>i, } with Nj pixels or superpixels Si
and a label set Y = {y\ , jj2--4jNi} with Nc labels, MRF
model constructs a posteriori probability of Su as shown in Eq.l.
Where, A > 0 is a constant coefficient. Pyi is prior probability
and Vij(yi,yj) = 1 when y r — y :i , Vij(yi,Vj) = 0 when
Ui Y y.j- Phi is always captured by feature-based discriminant
model like AdaBoost classifier and Pj i% tends to be
Puiiy-i\f(si)) where f{Si) is the features of Ay. For the whole
image, the posterior probability is P(Y\I, 0) where 0 is the
parameter of model. Some optimization algorithms, such as
GraphCut and Simulated Annealing Algorithm (SAA), can be
utilized to get maximums of P(Y\I, 0).
p {y, \s,)« p u {y, I s t ) P Vi (y, | y n )
Pn{y,\yvi) = tW\ ¿Z V v(yt>yj)
V y,en
P(Y\I,Q) = fjP(y,.\s i )
Y = argmax{P(71 /,©)}
As one of the most popular description model, MRF model can
balance the likelihood and prior probability in the whole image
and get global optimal solution with optimization algorithms
like algorithm presented in (Boykov Y, 2001). So, a linear
target prior can be introduced into MRF model in this paper
simply and obviously, see details in section 3.
3. METHODOLOGY
Once again, the main motivation of this paper is improving
region edges in classification results. So, the traditional edge
detection can get wealth and accurate edge information from
SAR images which is useful for classification. And, the
proposed approach in this paper gets use of this information in
over-segmentation, see details in section 3.
3.1 Pre-segmentations and Features Extraction
The proposed approach in this paper begins with pre
segmentation strategy using over-segmentation method to get
superpixels. Firstly, we utilize ratio edge detector to get edges
probability map of each input image (as shown in fig. 4.b).