IEEE - Aerospace and Electronic Systems - November 2021 - 11

Amit and Mohan
These region proposals also avoid specifying filters or
images of multiple scales or aspect ratios. In our network,
the RPN generates the region proposals using a sliding
window over the feature map to obtain features from CNN
(base network).
The extracted feature vector as shown in Figure 5 is
fed to two parallel network layers containing fully connected
layers for generating the bounding box coordinates
and prediction scores. For each pixel on the feature map,
we add a rectified linear unit (ReLU) activation function,
thus increasing the convergence rate and providing nonlinearity.
The final output of the first layer is the bounding
box coordinates/rectangular positions per anchor relative
to different scales (anchor scales) and aspect ratios
(anchor ratios). The second layer outputs the rectangular
position scores representing how likely the image in that
bounding box (positive anchors) will be a runway.
Nine region proposals known as anchor boxes are preFigure
8.
Generalized intersection over union (GIoU).
enable cross-layer connectivity and uninterrupted gradient
flow from the first layer to subsequent layers as shown in
Figure 6(a). The basic features in the primary layers are
fed to later layers, thus preserving the information, maintaining
spatial dimensionality between layers, and allowing
for fine-grained details to be classified better.
The maximum pooling layer is replaced by a 33 convolution
layer with stride-2. Thus, minimizing the number
ofparameters and preventing the loss of low-level features
attributed due to pooling. Images are resized to one size of
640640 pixels and the aspect ratio is preserved by zeropadding
at top/bottom or right/left for all nonsquare
images to support multiple images per batch. After extraction
of features using DarkNet-53, the original image is
transformed into a feature map of size 2020.
Techniques like upsampling, multiscale fusions for
multiscale feature extraction are implemented using the
FPN design, thus reducing the overall number of network
layers. Here, a pyramid of the same image is used at different
scales to detect objects. Generally, as we go up the
path, the spatial resolution decreases and thus semantic
values for each layer decreases. In our network, a topdown
approach is used wherein higher resolution layers
are created from a semantically rich layer. Object detection
is performed by feeding the extracted features to the
RPN.
The RPN as shown in Figure 7 identifies those regions
where the possibility of finding the runway (target object)
is high and proposes a subset window with an objectness
score. These networks share convolution layers with the
object detection networks and predict regions with multiple
scales as well as aspect ratios. Thus, maintaining high
recall rates and reducing the cost for computing proposals.
NOVEMBER 2021
dicted at three scales of 1282, 2562, and 5122 and three
aspect ratios of 1:1, 1:2, and 2:1 in the original faster
R-CNN algorithm [27]. However, it is observed that
images with a high spatial resolution (e.g., 10 m) consist
of many small runways, hence, requiring the inclusion of
two additional anchor scales, namely, 322 and 642 Also,
based on the elongated linear geometric shapes of the runway,
the aspect ratios 1:1, 1:2, and 2:1 cannot accurately
locate runways, hence two additional aspect ratios 1:3 and
3:1 are added in our implementations. Our framework utilizes
25 anchor boxes for each sliding window on the convolutional
feature map to determine the presence of more
than one runway in an image. The objectness scores at
each position, i.e., the probabilities of each prediction box
being a runway or background are calculated using a softmax
function and regression layer as shown in Figure 7.
The anchor labeling depends mainly on the values of
the intersection over union (IoU). The anchors are labeled
when IoU overlap is highest with ground truth and when
the IoU is higher than a given threshold. The threshold is
determined based on the complexity of the object to be
detected. The properties, such as the identity of indiscernible,
nonnegativity, symmetry, and the triangle inequality
are fulfilled by IoU as a distance metric. It is also scaleinvariant,
however, IoU suffers from either an approximation
or a plateau that exists in nonoverlapping cases.
In the context of the runway, demarcating single or
multiple runways (parallel runway, intersected runways,
V-runways) become difficult using IoU, as it cannot reflect
the overlap situation for two rectangular boxes (A and B).
Also, in scenarios when IoUðA;BÞ¼ 0, it is impossible
to decipher if A and B are adjacent to each other or far
apart. Hence, generalized intersection over union
(GIoU) [28] as shown in Figure 8 is applied here to measure
the similarity between ground-truth and the prediction
(anchor) bounding boxes, and between different
prediction boxes. Generalized intersection over union is
IEEE A&E SYSTEMS MAGAZINE
11
IEEE - Aerospace and Electronic Systems - November 2021

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