![](https://csdnimg.cn/release/download_crawler_static/10379776/bg4.jpg)
Remote Sens. 2017, 9, 439 4 of 19
2.1. Line Segment Detection
Segment fragmentation is a common problem for line detection algorithms, and it has a significant
influence on the results of image registration. If the scales of the reference and test images are
estimated, then we can unify the scales of these two images via downsampling to overcome the
segment fragmentation problem. Thus, we propose a strategy for unifying the scales of images via an
image pyramid. First, line segments are extracted from the image pyramid. Second, local descriptors
for line segments from different scales are generated. Third, the scales of the two images are estimated
by calculating the Euclidean distances of the descriptors. Lastly, the scales of the two images are unified
into the same level, and the line segments are matched based on the distances of their descriptors.
The image pyramid line segment detection method is presented in this section.
In this study, we assume that all the input images contain a considerable amount of artificial objects
(e.g., roads, buildings, farmlands, and airports) that can be detected as line segments. To overcome the
problem of segment fragmentation, an image pyramid that consists of n octaves is used. In this image
pyramid, each octave is generated by downsampling the original input image with a series of scales,
and the first octave is established as the input image. Furthermore, each octave has only one layer;
such a scenario differs from the traditional Gaussian pyramid. We can obtain the multi-resolution
representation of an image by using this image pyramid.
To detect line segments, EDLines [
35
] algorithm is used in this paper. EDLines is an automatic
line segment detection algorithm with no requirement for parameter tuning, and it is one of the
state-of-the-art line detection algorithms [
36
]. It can obtain meaningful and accurate segments by
using a false positive control criterion named the Helmholtz principle [
37
]. Thus, EDLines is suitable
for extracting line segments from the main shape contours in remote sensing images. As shown in
Figure 2, EDLines algorithm is applied to each octave image in the image pyramid.
...
Scale
LineScale[1]={ }
LineScale[2]={ }
LineScale[3]={ }
LineScale[4]={ }
LineScale[n]={ }
Figure 2.
Illustration of line segment detection in an image pyramid. The first octave is the input image,
and the other octave images are downsampled from it. Line segments detected from the ith octave
image are represented as LineScale[i].
Although EDLines algorithm can obtain meaningful segments as mentioned above, there still
exists segment fragmentation in practice. In this study, a line-merging procedure is performed
according to the parallel relation and distance of the line segments. The underlying idea of line-merging
procedure is to merge two segments that lie on one line belonging to the same shape contour. This
procedure can reduce the number of segments by an average of 20%, thereby not only helping improve
segment quality but also reduce computation cost. It is worth mentioning that the proposed RLI can
work without this line-merging procedure. Then, all the remaining line segments from the ith octave