IEEE - Aerospace and Electronic Systems - November 2021 - 8

A Robust Airport Runway Detection Network Based on R-CNN Using Remote Sensing Images
Figure 4.
Generic Aerial vehicle navigation and guidance system architecture.
topology of convolutional neural network (CNN) makes
the model ideal for use in different resource-constraint
environments that require high accuracies, such as autonomous
vehicles, robotics, and healthcare.
Object detection combines feature extraction, image
classification, and localization by drawing one or more
bounding boxes around the extent of the objects as well as
predicting the type or class ofan object.
The backbone networks used for feature extraction
from input images or videos by the majority of the object
detectors are CNN architectures. A separate detection network
is developed for classification and localization tasks.
For example, the early convolutional neural network
(CNN) LeNet-5 consists of three layers: namely, a convolution
(CONV) layer, average pooling (AVG POOL)
layer, and a fully connected (FC) layer.
With the increased complexity of data and the need for
better computing power, deep and wide versions of CNN
architectures were introduced. The AlexNet [18] network
paved way for use of graphic processing units, thus
enabling 10 faster training time and allowing the use of
larger and bigger datasets like ImageNet. Other architectures,
such as VGG-16 [30], ResNet and its variants [14],
etc., have been used extensively for feature extraction.
These models introduced techniques such as dropout,
overlapping max pooling, network-in-network, and bottleneck
layers, etc., that provide a rich combination of features
with low computation time. ResNet architecture is
seen as ensembles of both parallel and serial modules,
which operate on blocks of relatively low depth.
8
The evolution of the region-based convolutional neural
network (R-CNN) [11] marked the beginning of the
widespread implementation ofCNNs to the various object
localization, detection, and segmentation issues. For
example, a most generic R-CNN network consists of a
RPN that generates and extracts candidate bounding boxes
using a selective search algorithm, followed by a feature
extractor. Feature classification to one of the classes is
finally done using classifiers, such as linear support vector
machine (SVM). Fast R-CNN [10] uses a fixed-length feature
vector extraction algorithm on a region of interest
(RoI) pooling layer. Region proposals from the same
image can thus share memory and computation in both
the forward and backward passes. Faster R-CNN [27], an
advancement in Fast R-CNN, popularly known as the
R-CNN model family, replaced the selective search
proposals with a RPN. In both these methods, the proposals
are learned from a CNN to simultaneously predict
object bounds and the objectness/probabilities scores at
each position.
On the other hand, the you-only-look-once (YOLOV3)
[24] model family introduced the single-stage detectors capable
ofmaking predictions directly using the grids. Here, the
components of the two-stage detectors are integrated into a
single neural network, and the object detection task is treated
as a simple regression problem. The class probabilities and
the bounding box coordinates for all classes are learned
simultaneously along with the input image.
A lightweight base model called DarkNet-19 comprising
of passthrough layers similar to identity blocks in ResNet is
IEEE A&E SYSTEMS MAGAZINE
NOVEMBER 2021
IEEE - Aerospace and Electronic Systems - November 2021

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