APPLICATION OF DEEP LEARNING IN OBJECT DETECTION

Abstract

In the field of object identification, a deep learning neural
network is a sort of neural network that has been widely utilised
since it was originally developed. An in-depth summary of the
project "Object Recognition by Deep Learning Neural
Networks" will be offered with the aid of this draught
preliminary interim report, which is currently under
development. After everything is said and done, the project's goal
is to: 1) replicate CNN on Python; and 2) replace the original
classifier with a Latent Dirichlet Allocation classifier in order to
increase the accuracy of findings. In order to get the intended
results, it will be essential to train and evaluate the algorithm
using publically available datasets, which will be made available
by the project team. The version CNN algorithm has been
developed as of the current stage of development, and selective
search has been copied in the Python version CNN algorithm,
which is currently at the forefront of development. So yet, there
have been no significant roadblocks encountered as a result of
the project's early development stage. We anticipate that the
finished product will be able to attain a greater accuracy rate
than the CNN implementation that was created in Python in the
first place. Computer vision problems such as the detection of
visual objects have been extensively researched, with deep
learning applied to real-world data being the primary method of
investigation. However, data obtained from virtual environments
has not gotten the same level of attention that data obtained from
traditional sources has received. When data is collected in a
virtual environment, it is possible to gather information from
sites that are not easily accessible for data collection on the
ground, such as aerial settings that are difficult to reach. As part
of a wider project, we are investigating how to recognise items
when they are placed in a digital context, particularly an aerial
virtual world. This is a portion of our research. For this
experiment, a simulator is utilised to build a synthetic data set
that contains 16 distinct types of automobiles that were recorded
from the air by an aeroplane and placed on a test track. We have
trained and evaluated two state-of-the-art detectors on the basis
of the data set that has been created, with the purpose of
evaluating the performance of current approaches in virtual
settings. Both the You Only Look Once version 3 (YOLOv3) and
the Single Shot Multibox Detector (SSD) achieve performance
quality that is comparable to that stated in the literature,
according to the findings of real-world testing conducted by the
authors. Also examined are several types of fusion processes
between detectors that have been trained on two different subsets
of the same dataset, for example, a subset of vehicles with
constant colours and an entirely different dataset of automobiles
with changeable colours. Because of their trials' findings and
subsequent analysis, the researchers realised that it is feasible to
train many instances of the detector on distinct subsets of the
data set and then combine these detectors in order to improve the
performance.