Aerospace and Electronic Systems - May 2019 - 60

Power Line Detection for Aircraft Safety Based on Image Processing Techniques: Advances and Recommendations

Figure 10.
PLD based on multimodal deep learning.

PLD BASED ON MULTIMODAL LEARNING
The multimodal learning can be employed to effectively
process information of different modalities simultaneously.
As illustrated in Figure 10, effective features learned in
different models can be fused to get better detection
performance.
The strength of this method is that effective
features in different information, such as visible and
infrared images, can be efficiently learned and automatically extracted. Note that the deep learning model
1 and model 2 in Figure 10 can be different. This is
because that information with different modalities
have different characteristics, so the suitable network
for them may vary.

To evaluate the performance of PLD algorithms, ground
truth data is necessary. Besides, evaluation metrics would
vary when formulating PLD as a detection and classification problem.

PLD AS A CLASSIFICATION PROBLEM
In this case, two metrics namely precision and recall could
be chosen, which are defines as
true positive
true positive + false positive
true positive
recall ¼
true positive + false negative

(1)

where the true positive means the percentage of examples
which contain power lines are identified as having power
cables. False positive means the number of examples
which are predicted as having power lines but actually
not. False negative means the percentage of examples
which are predicted as not containing power lines but
60

PLD AS A DETECTION PROBLEM
In this case, the intersection-over-union (IoU) can be
employed to measure the performance of PLD algorithms.
Specifically, the IoU is defined as the overlapping rate of
candidate bound and ground truth bound, namely:
IoU ¼

T
areaðCÞ areaðGÞ
areaðCÞ [ areaðGÞ

(2)

where areaðCÞ and areaðGÞ are the area of candidate
bound and ground truth bound, respectively.

QUANTITATIVE MEASURES OF PERFORMANCE

precision ¼

actually do. After obtaining precision and recall, one can
draw the P-R curve whose y-axis is precision and x-axis is
recall, respectively. The area under curve in P-R plot can
then be utilized to measure the overall performance
of PLD algorithms.

CONCLUSION
In this paper, an extensive survey on PLD methods for
low-altitude aircraft safety based on image processing is
presented. First, primary features of power lines are
introduced. Then, main studies on PLD are summarized,
including PLD approaches based on tradition image processing methods and machine learning. Besides, a few latest research works, such as PLD based on deep learning
and infrared images, are discussed.
By discussing and comparing current methods, we have
proposed future development directions of PLD, with the
aim of providing a starting point for researchers in
PLD. Specifically, image fusion and multimodal sensor
fusion should be able to provide better detection performance and lower false alarm rate by exploiting complementary information from different sensors. In addition, AI will
be an important technique in PLD, with the help of transfer

IEEE A&E SYSTEMS MAGAZINE

MAY 2019

Aerospace and Electronic Systems - May 2019

Table of Contents for the Digital Edition of Aerospace and Electronic Systems - May 2019