IEEE - Aerospace and Electronic Systems - November 2021 - 12

A Robust Airport Runway Detection Network Based on R-CNN Using Remote Sensing Images
Figure 9.
Different phases of candidate proposal refinement: Before the application of RoI pooling layer, before and after the non-maximum suppression,
and the final detection results.
defined as follows:
GIoU ¼ IoU
jCnðA [ BÞj
jCj
(1)
where IoU is used to evaluate how close the prediction
bounding box is to the ground truth, A and B are the predictions
and ground truth boxes, respectively. The variable
C is the smallest convex hull that encloses both A and B
and is known as the minimum enclosing bounding box of
A and B.
From (1), it is seen that a smaller value is subtracted
from IoU, thus increasing the value ofGIoU as the prediction
moves toward the ground truth, maintaining both the
scale-invariance and distance metrics properties. Also,
GIoU is less than or equal to IoU, which reflects a better
overlap between the bounding boxes. It is also observed
that IoU has no value when there is no intersection, and
therefore no gradient, but GIoU is always differentiable.
The loss function must be differentiable to allow for backpropagation
and, hence, GIoU is used as the regression
loss (distance loss) function for the prediction box. The
regression loss function is defined as in (2) and (3):
Lreg ¼ 1 GIoU
where
GIoU ¼
1 when A = B, GIoU ¼ IoU ¼ 1
1 when A \ B ¼; and is far away.
(3)
An anchor is labeled as runway or background based on
the GIoU ratio as defined as:
An anchor is labeled positive if the ratio ofGIoU to
ground truth is highest in case of multiple ground
truths, or greater than 0.7, and
If the GIoU ratio is less than 0.3, then the anchor is
labeled negative (background).
The rest of the anchors that do not meet the above criteria
are not used while training, therefore lowering the
process overhead. The features bound by the region proposals
are called the region of interest (RoI). Using the
12
IEEE A&E SYSTEMS MAGAZINE
Lðfpig; ftigÞ ¼
1
Ncls
þ
X
i
1
Nreg
Lclsðpi;p
X
i
p
i Þ
(5)
i Lregðti;t
i Þ
(2)
maximum pooling operation, the RoI pooling layer as
shown in Figure 9 takes the region proposals with varied
sizes to generate a fixed size of the input (fixed-size feature
vector) for the next layer, which is a fully connected
layer. Regression and softmax layers are included after
the fully connected layer to predict the bounding boxes
and the classification scores. The optimal candidate box is
generated using a cascade RPN with soft-decision nonmaximum
suppression (NMS) that reduces the number of
candidate boxes. Thus, reducing the missed runway detection
while suppressing the redundant detection boxes.
LOSS FUNCTION FOR LEARNING
The loss function for the ARDN network is determined
using both the classification loss and bounding box regression
loss, which is similar to the fast R-CNN [10] network.
The network is hence optimized for a multitask loss function
as in (4) and (5). During the training phase, a binary
class label is assigned to each anchor to indicate if the
anchor is a runway or background (scenes with no runway).
The anchor labeling (positive or negative) is determined
based on how close or far the GIoU value is from
the predefined threshold.
The classification loss Lcls is the log loss of two
classes (runway, background) and is as defined in (6).
To improve the loss function of the prediction center point
coordinate (X;Y), generalized intersection over union
(GIoU) is applied.
For the regression loss Lreg, Lregðti;t
i Þ¼ Rsðti t
i Þ
is applied, where Rs is the smooth loss as in (7). The regression
loss is activated only for the positive anchors, and not
otherwise
L¼ Lcls þLreg
(4)
NOVEMBER 2021
IEEE - Aerospace and Electronic Systems - November 2021

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - November 2021
