05-09-2016, 09:23 AM
A Novel Airport Detection Method via Line Segment
Classification and Texture Classification
1452582068-airportdetection.pdf (Size: 1.15 MB / Downloads: 8)
Abstract—Airports are one of the most important traffic facilities;
thus, airport detection is of great significance in economic
and military construction. This letter proposes a novel method
for airport detection, with the entire algorithm based on line
segment classification and texture classification. First, a fast line
segment detector is applied to extract the line segments in images
and compute the features of these line segments. Then, the line
segments are discriminated by a trained runway line classifier, and
the regions of interest (ROIs) are extracted from the line segments,
which are classified as runway lines. Finally, whether the ROI is
actually an airport is determined by analyzing the classification
results of the image blocks. This method is unique in terms of the
computing of line segment features and line segment classification.
Experimental results demonstrate the effectiveness and robustness
of the proposed method.
Index Terms—Airport detection, line segment classification, line
segment detector (LSD), region of interest (ROI), support vector
machine (SVM), texture classification.
INTRODUCTION
AS ONE of the most important traffic facilities, airports
have played a critical role in economic and military construction.
Detecting airports automatically from remote sensing
images has great practical value for aircraft autopilot operations,
airport navigation, and many other fields. It has been an
important research topic in object recognition.
In recent years, many scholars have studied this issue, and
multiple airport detection methods have been proposed. These
methods can be divided into two categories. The first category is
detecting the location of the runways based on their geometric
structure [1]–[6]. Runways often have parallel straight lines and
special lengths and widths; thus, they can be identified based
on this knowledge. This approach is a decision tree method
based on prior knowledge; it requires many criteria, and each
criterion plays an important role in the result. One inappropriate
criterion can lead to an inaccurate result. Furthermore, the method only focuses on the shape of the runways, which can be
confused with the edges of roads, bridges, and shorelines. The
other detection method is texture classification [7]–[11]. Image
blocks are classified after texture feature extraction, and the
airports are determined according to the classification results.
The texture classifier is generally trained by kernel matching
pursuits [7], support vector machine (SVM) [8], and Adaboost
learning [9]. In contrast, in [10], the regions of interest (ROIs)
are extracted by exploiting the clustering information from
matched SIFT key points, and in [11], the visual saliency model
is applied to compute the saliency map and extract ROIs. This
classification method is more intelligent than the knowledgebased
approach. However, it is time-consuming if the entire
blocks of images require classification. Therefore, the first
method is typically used to extract the ROIs, and then, each
ROI is verified to determine whether it is an airport via texture
classification.
Overall, airport detection research has advanced considerably
in recent years. However, it still requires improvement,
particularly when considering high-resolution and complex
background images. In this letter, a novel airport detection
method via line segment classification and texture classification
is proposed. The method consists of double classifiers and
four main sections. First, some features of runway lines are
learned by the SVM [12] to obtain the runway line classifier,
and the texture features of the runway blocks are also learned
by the SVM to obtain the runway block classifier. Second,
a fast line segment detector (LSD) [13] is applied to extract
the line segments in images and compute the features of these
line segments. Third, the line segments are discriminated by
the runway line classifier. Moreover, the ROIs are extracted
from the line segments, which are classified as runway lines.
Finally, each ROI is divided into blocks of a specific size and
classified by the runway block classifier. Then, the classification
results are judged to determine whether the ROI is really an
airport. Experimental results on satellite images demonstrate
the effectiveness and robustness of the proposed method.
II. LINE SEGMENT CLASSIFICATION
For airports to be automatically detected from the images,
the regions that may contain an airport must first be determined;
these regions are called ROIs. In this letter, an intelligent
ROI extraction method based on line segment classification
is proposed. By analyzing the line segment features, an SVM
can be used to build a suitable classifier that can effectively
distinguish runway lines from the other landforms.
A. Line Segment Detection
LSD [13] is an excellent line segment detection method that
can yield accurate subpixel results in linear time. It is based
on phase grouping and designed to work on any image without
parameter tuning. In the algorithm, a line segment is defined
as the local straight contours whose pixels share the same
gradient orientation. As shown in Fig. 1, LSD can obtain almost
all of the runways and other line segments at an airport; the
result includes the endpoints and width of each line segment.
However, the detected line segments corresponding to the long
linear edges in the image are generally not continuous because
each point can only belong to one line segment in the algorithm.
B. Line Segment Feature Extraction
It is difficult to determine which line segment is the runway
line based solely on the endpoints and width. More detailed
information is required, including the relationship between the
line segments. Moreover, because the runways are parallel to
each other, only the parallel line segments need to be considered.
Thus, the line segments were grouped by their angles,
and the collinear line segments in each group were linked. As
illustrated in Fig. 2(a), the original line segments are within
a group because they share approximately the same angle.
Fig. 2(b) is obtained after linking the collinear line segments
in Fig. 2(a). Then, according to the performance of LSD and
the characteristics of the airport runways, the following information
is gathered as the features of each original line segment.