Abstract
The detection of targets under the ground is an important procedure
that is typically performed by a human non-automatically. Recent studies
have automated this process using artificial intelligence (AI) based on
radar images. There are three main steps before feeding reconstructed
radar images to a neural network. The first step is segmentation which
can make the detection task straightforward. We have proposed an
Otsu-based segmentation algorithm in this paper. The proposed
segmentation algorithm is effectively able to distinguish between all
the targets. In the second step before employing AI to detect targets, a
local sliding window has been taken into consideration to improve the
results. The image is divided into smaller parts by this sliding window
after it has been reconstructed. In the third step, two different
methods have been considered for data augmentation. The first method is
a novel approach for generating synthetic radar data. It is applied
before radar image reconstruction based on the summation of two
receivers’ signals with different coefficients. In the second
augmentation method, some conventional data augmentation methods like
flip and rotating are applied to complete this task. To discriminate
targets from background, it is necessary to classify input images to
AI-based aproaches. This task can be accomplished by classical machine
learning approaches like the scalar vector machine (SVM). Gabor filters
have been utilized in this paper to extract the features. There also
exist two classification approaches using convolutional neural networks
(CNN) to automatically detect targets after image reconstruction. Two
different CNN have been implemented. Without data augmentation, the
SVM-based approach works better than CNN, and its accuracy is 86.9%.
Overall, the second CNN algorithm outperformed SVM after the data
augmentation by reaching 96% accuracy.