IEEE Geoscience and Remote Sensing Magazine - December 2015 - 25

C. Gray histoGram
The spectrum of the original image can also show diversity. The poor texture makes the pixel-to-pixel comparison
ambiguous, and the perspective or illumination changes
may affect the gray values. Nonetheless, the region histogram considers that all pixels and some similarity measurements can overcome these problems. In this study, a
fast similarity measurement, i.e., histogram intersection
[18], is adopted to select correspondence. According to the
gray histograms (H(i)) of building roofs, this measurement
is calculated as follows:
s H = 1- / i min (H 1 (i), H 2 (i))

(3)

III. EXPERIMENTS AND COMPARISON
The proposed method was applied over an area in Yangjiang, China. The images are captured by the SWDC-5 looking nadir with a 60% overlap, thus supplying favorable
data for image matching. The ground resolution of images
is close to 10 cm when the flying height is about 850 m,
which is clear enough to detect building roofs. A test area
(about 5000 # 3000 pixels) with similar building roofs is
taken to evaluate the proposed matching approach in complex urban scenes. An application case study on an overlapping area is conducted for image mosaicking to verify the
matching validity. Fig. 4 shows the experimental images.
3.1 Results of the Regional
featuRe constRuction
For each image, line detector and K-Means is preprocessed preparing for the external contour and internal

structure. The line segments are extracted thoroughly
by the EDLines detector (Fig. 5). Although the fragmentation is remarkable in the roof contour, the edge
feature is fully detected such that it is feasible to restore the straight line by reorganization. The classified
map (Fig. 6) shows the properties of a simple structure,
quick response, and robustness.
Fig. 7 represents the regional features constructed by
combining the contour and internal blocks. The adopted
building extraction algorithm can fully restore the contour,
and the classification-based approach can robustly obtain
simple heterogeneous distribution, thus providing stable
and distinctive features for
AN APPlICATION
image matching.
CASE STuDy ON AN

3.2 evaluation
OvERlAPPINg AREA IS
of matching
CONDuCTED FOR IMAgE
Image matching is conducted
MOSAICkINg TO vERIFy ThE
on the basis of the feature deMATChINg vAlIDITy.
scription and similarity measurement, following the extraction of regional features.
To represent the dispositions of the 55 roofs in the overlapping area, we use a common confusion matrix (Table 1)
to evaluate the true and false matches and match failures
[19]. TP, FN, FP, and TN represent the number of correct
matches, missed matches, incorrect matches, and nonmatches, respectively, all of which were correctly rejected.
TPR, FPR, PPV, and ACC are unit rates that quantitatively
describe the good matching results. The negative matching results mainly come from missed matches instead

FIguRE 4. The overall images are applied to experimentally image

mosaicing, and the similar areas within green boxes are used to
specifically evaluate the matching algorithm.

FIguRE 6. The classified map processed by K-Mean.

FIguRE 7. Regional features constructed of internal structure and
FIguRE 5. The line segments extracted by EDLines.
december 2015

ieee Geoscience and remote sensing magazine

contour.
25

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2015