sensors

Article
Bangladeshi Native Vehicle Classification Based on Transfer
Learning with Deep Convolutional Neural Network
Md Mahibul Hasan 1 , Zhijie Wang 1, *, Muhammad Ather Iqbal Hussain 1 and Kaniz Fatima 2

1 College of Information Science and Technology, Donghua University, Shanghai 201620, China;
[email protected] (M.M.H.); [email protected] (M.A.I.H.)
2 Institute of Business Administration, Jahangirnagar University, Savar, Dhaka 1342, Bangladesh;
[email protected]
* Correspondence: [email protected]

Abstract: Vehicle type classification plays an essential role in developing an intelligent transportation
system (ITS). Based on the modern accomplishments of deep learning (DL) on image classification,
we proposed a model based on transfer learning, incorporating data augmentation, for the recog-
nition and classification of Bangladeshi native vehicle types. An extensive dataset of Bangladeshi
native vehicles, encompassing 10,440 images, was developed. Here, the images are categorized into
13 common vehicle classes in Bangladesh. The method utilized was a residual network (ResNet-50)-
based model, with extra classification blocks added to improve performance. Here, vehicle type
features were automatically extracted and categorized. While conducting the analysis, a variety of
metrics was used for the evaluation, including accuracy, precision, recall, and F1 − Score. In spite of





 the changing physical properties of the vehicles, the proposed model achieved progressive accuracy.
Citation: Hasan, M.M.; Wang, Z.;
Our proposed method surpasses the existing baseline method as well as two pre-trained DL ap-
Hussain, M.A.I.; Fatima, K. proaches, AlexNet and VGG-16. Based on result comparisons, we have seen that, in the classification
Bangladeshi Native Vehicle of Bangladeshi native vehicle types, our suggested ResNet-50 pre-trained model achieves an accuracy
Classification Based on Transfer of 98.00%.
Learning with Deep Convolutional
Neural Network. Sensors 2021, 21, Keywords: native vehicle type classification; Deshi-BD vehicle dataset; deep learning; transfer
7545. https://doi.org/10.3390/ learning; ResNet-50
s21227545

Academic Editors: Javier Alonso Ruiz,

Angel Llamazares and Martin Lauer 1. Introduction
Road traffic accidents are a global concern due to the increasing amount of people
Received: 12 October 2021
Accepted: 5 November 2021
who die, or are extremely injured, because of these accidents. Statistics show that, each
Published: 13 November 2021
year, around 1.2 million people die as a result of road incidents. Moreover, statistics
show that over 50 million people are injured in road accidents globally [1]. However, the
Publisher’s Note: MDPI stays neutral
phenomenon varies by country. When compared with developed countries, the number of
with regard to jurisdictional claims in
injuries, deaths, and accidents are 10 to 60 times higher in developing nations [2]. Over 80%
published maps and institutional affil- of total road injuries take place in the world’s developing regions [3]. Bangladesh is one of
iations. a few developing nations where the rate of injuries, deaths, and accidents is extremely high.
The current situation is worse than ever before. For example, 20 people on average die
each day because of road accidents. As per the guidelines from the United Nations Road
Safety Action Plan 2011–2020, the Sustainable Development Goals (SDGs) 2030, and the
Copyright: © 2021 by the authors.
associated GOAL-3.6, Bangladesh is required to cut the number of road traffic injuries and
Licensee MDPI, Basel, Switzerland.
deaths in half [3,4]. Considering this and keeping in harmony with developed countries,
This article is an open access article
it is crucial for Bangladesh to be dependent on an intelligent transportation system, to
distributed under the terms and develop its traffic management system.
conditions of the Creative Commons An intelligent transportation system (ITS) is a widely used term, related to the con-
Attribution (CC BY) license (https:// cept used in road and transportation planning. The aim is to enhance the performance
creativecommons.org/licenses/by/ and security of, for example, superhighway tolls, traffic counts, and traffic observations.
4.0/). One outstanding feature of an ITS is vehicle type classification. It offers a broad range of

Sensors 2021, 21, 7545. https://doi.org/10.3390/s21227545 https://www.mdpi.com/journal/sensors

Sensors 2021, 21, 7545 2 of 21

applications, including smart parking solutions, traffic management statistics, and identifi-
cation of vehicle types. Prevailing strategies rely on ultrasonic, sensors, and video devices.
The computer vision (CV) community is focusing on image-based methods due to the
extensive utilization of vehicle monitoring devices. From the perspective of an ITS, object
classification of vehicles plays an essential role. It has an extensive variety of engagement,
consisting of traffic monitoring, routing, and tracking. Object classification involves an
enormous discipline of research in regard to image processing techniques and it seeks
to classify elements in images, into significant groups [5]. Individuals can categorize ve-
hicles, without difficulty, from images via key aspects, such as trademarks, forms, and
ornaments. Nonetheless, classification of vehicle types in images is, perhaps, a tough
issue for computer systems. This is due to the fact that image sources have multi-scale
characteristics [5]. Vehicles are also available in a wide range of shapes, measurements, and
shades. Furthermore, natural factors, such as illumination, noise, complex background,
and climate, affect the ability to capture photos in traffic.
For vehicle type classification [6–9], techniques based on laser and loop induction
detectors have already been presented. In these methods, with the aim to collect and
analyze records and bring out relevant information regarding vehicles, the sensors were
installed under road pavements. However, adverse weather and damage in walkways
are responsible for influencing the correctness and stability of these methodologies [10].
CV has progressed tremendously in recent years. Thus, the authors have suggested
the use of vehicle classification systems, which are based on pattern identification and
image analysis [11,12]. This is actually a process consisting of two stages. The first stage
incorporates visual characteristics from input visual frames using handcrafted extraction
methods. In the second stage, training on the extracted features is provided to the machine
learning (ML) classifiers, in order to categorize data, depending on the types. Moreover,
there are two types of customized characteristics: (i) global and (ii) local. The division helps
to define and categorize the image information [13]. All of these attributes are used to train
existing ML classifiers. Consequently, they help with object classification. The mentioned
methods perform very well in specific regulated surroundings. Moreover, these systems
are more convenient to install (and look after) as compared to existing methods, which are
dependent on laser and inductive processes. However, the algorithms are given training
on insufficient customized characteristics derived from limited datasets, but maintaining
accuracy in a temporal setting requires significant prior information [14].
Deep learning (DL)-based feature extraction and classification approaches have re-
cently gained popularity, exhibiting superior adaptability and flexibility over conventional
methods. Because of their improved design, CNN classification algorithms gained notable
precision on wide-range databases [15–17]. To date, as far as we know, for the creation and
application of Bangladeshi vehicle classification systems, there is no universal benchmark
dataset available. The existing vehicle classification datasets, such as CompCars [18] and
Stanford Cars dataset [19], are very small. These are based on narrow classifications of
certain locations. Encouraged by the prior studies and the enthusiasm to address the
limitations, this paper will focus on the vehicle type classification on roads, so that various
nations, especially South Asian nations (e.g., Bangladesh), could take advantage from its
implementation. These countries are still using conventional strategies manually observed
via human, photography, and sensor-based programs. Thus, a competent approach of the
traffic surveillance system in Bangladesh is required to acquire accurate final results. By
applying data augmentation and transfer learning approaches, we present a DL model for
recognition and classification of Bangladeshi native vehicles.
To overcome all of the above problems in Bangladeshi native vehicle classification
systems, we established the below-mentioned improvements to our native vehicle classifi-
cation system.

(i). We introduced a Bangladeshi native vehicle dataset, the Deshi-BD dataset, which
contains 10,440 images based on 13 Bangladeshi on-road vehicle classes. These images
were manually collected from driving videos by us. It is important to note that these
Sensors 2021, 21, 7545 3 of 21

classes are distinct, in terms of features and shapes, and they are not limited in the
current datasets.
(ii). A pre-trained CNN model using ResNet-50 [20] was implemented to increase the flex-
ibility of Bangladeshi native vehicle classification techniques under poor illuminating
circumstances.
(iii). We evaluated various performance measures for our native vehicle classification
model, including (i) accuracy; (ii) precision; (iii) recall; and (iv) F1 − Score. We also
compared our suggested approach with AlexNet and VGG-16 CNN models that have
been trained.
(iv). Finally, an overall comparison of the anticipated and standing vehicle classification
techniques were performed to present an accurate possibility of our native vehicle
classification network proposal. As seen by the results, our suggested technique
provides greater classification precision when compared to other traditional methods.
The following is how we structured our paper. We present several similar studies
and emphasize their findings in Section 2. We explain the proposed approach in Section 3.
We provide the analysis of results in Section 4. In Section 5, we offer a discussion and
a comparison with previous works. Finally, in Section 6, we provide a conclusion and
suggestions for further work.

2. Related Work
Artificial intelligence is developing at a high speed. Accordingly, vision-based ve-
hicle classification is regarded as a critical component of driverless vehicle observation
units. The two primary forms of vision-based vehicle classification techniques are the
(i) customized feature-based technique and (ii) deep feature-based technique [10]. These
groupings are found in existing research works. During the primary period of CV, cus-
tomized feature-based approaches were expected for ITSs. Researchers used the HOG-SVM
based customized features approach for training a SVM classifier utilizing HOG features
along with the Gaussian Kernel feature suggested by Ng et al. [21]. The aforementioned
classification model was tested on a surveillance footage collection of 2800 images. It
classified three types of vehicles with 92.3% correctness. Wen et al. [22] conducted the
study, applying an AdaBoost-based rapid learning vehicle predictor to separate data, which
were categorized as (i) non-vehicle and (ii) vehicle.
Furthermore, for the rapid learning of classifiers, the authors suggested a procedure
for generating Haar-like attributes. This current classifier was tested for an open Caltech
data source and gained correctness of 92.89%. Matos et al. [23] presented a combined
method for integrating vehicle characteristics such as width, height, enclosing lines, etc.
These were based on two neural networks. This suggested architecture, having a sample
size of 100, accomplished 69% correctness. On the other hand, a classification method
was demonstrated by Chen et al. [24] that extracted both texture and HOG attributes.
It also classified the vehicles by utilizing a fuzzy enthused SVM classification model.
The demonstrated classifier achieved accuracy of 92.6%. It was evaluated on a dataset
having 2000 images of cars, vans, and buses. A collective method, integrating the SIFT
classifier and BoW-based method, was suggested by Cui et al. [25] with the aim to extract
the characteristics and apply SVM to classify the dataset. The dataset was a group of
340 images of trucks, cars, and minibuses. The ultimate result achieved a 90.2% accuracy
by the presented classifier.
Moreover, deep feature-based systems can mitigate the issues related to handcrafted
feature-based classifiers. A CNN-based semi supervised classification system for simul-
taneous vehicle classification was demonstrated by Dong et al. [26]. In their work, a
sparse-Laplacian filter-dependent process was planned to extract comparative vehicle
data. Moreover, to compute the class likelihood of the associated vehicle, a softmax layer
was trained. They evaluated the data using the Bit-Vehicle database, where day scenes
demonstrated 96.1% correctness and night scenes demonstrated 89.6% correctness. For
vehicle sorting in an uncontrolled road atmosphere, a CNN and an end-to-end combined
Sensors 2021, 21, 7545 4 of 21

model were suggested by Cao et al. [27]. The expected structure succeeded in providing a
95.3% correctness, evaluated on the CompCars view-aware dataset.
Moreover, Jo et al. [28] established GoogLeNet architecture that focused on transfer
learning. It was used for vehicle classification of road traffic; the classifier achieved 98.3%
accuracy. The research was based on the ILSVRC-2012 dataset. The classification and
identification of vehicles on highway routes has a strong influence on traffic and accident
management. CNN architecture, which focuses on the vehicle classification algorithm for
vehicle classification and numbering on major routes, was an idea from Chauhan et al. [29].
After applying 5562 CCTV camera videos on highways, they demanded that their offered
model achieved 75% MAP. In the research work by Kim et al. [30], they used the PCANeT-
HOG-HU model that focuses on the collective characteristic extracting procedure. In this
case, the methodology was fed into SVM as input data for training the classifier architecture.
The authors extracted 13,700 images of cars from surveillance recordings for training and
testing the suggested classifiers model. The dataset comprised six types of vehicles. They
suggested mild-mass classification architecture, which obtained a mean correctness of
98.34%. Fast R–CNN architecture, which focuses on vehicle classification techniques, was
suggested by Wang et al. [31]. The aim was to develop a technique for traffic surveillance
in a real-time atmosphere. The authors collected and tested a sample with 60,000 images.
These data were gathered and separated to train and test sets. The total correct result
obtained was 80.051%.
Other researchers have proposed a method, focusing on deep features, which could
significantly improve vehicle recognition accuracy. However, they require a large amount of
data to achieve considerable precision in the present ITS operations [32–35]. In the modern
era, extensive research has been conducted in this field; yet, the current civic resources for
vehicles or ITSs include automotive kinds. These are common in wealthy nations. The
classification methods, however, are unrealistic regarding ITSs in South Asian nations. All
identified difficulties point to the need for a unique vehicle classification methodology, as
well as a collection that includes common vehicles, such as conventional buses, trucks, cars,
CNG, motorbikes, rickshaws, auto rickshaws, and vans in South Asian nations.

3. Materials and Methods

To solve the above-mentioned problems, we propose a new vehicle dataset consisting
of 10,440 images of Bangladeshi typical traffic vehicles, divided into thirteen types. To
improve the performance of the recommended classification in real-time ITS applications,
the ultimate system was customized using existing AlexNet [36], VGG [37], and ResNet [20]
models. Depending on the performance of these models, the top performing model was
implemented to improve the accuracy of our system. As a result, we saw that ResNet
surpasses the other models in terms of closure, accuracy, and processing speed. Therefore,
the ResNet model with 50 layers was improved and implemented in our proposed vehicle
classification system. The proposed method is described in detail in Figure 1.

3.1. Deshi-BD Vehicle Dataset

The data source is an important component in supporting algorithms in learning
features and making predictions based on the learned information when using DL based
classification systems. To the best of our knowledge, there seems to be no standardized
Bangladeshi general vehicle dataset that includes data on Bangladeshi native vehicles and
that solves classification problems. As documented, CompCars as well as the Stanford
vehicle database simply incorporate the types of modern vehicles in specific areas. These
cannot be used in actual-time classifiers in other geographies, such as Bangladesh. In
addition, the recommended data source differs from the available data sources. There is a
lot of variation regarding characteristics and structure. Furthermore, the present vehicle
classifier may not execute well enough in practical ITS implementations, as it is developed
on fairly short data samples with few classes [38]. To address these issues, we created
Sensors 2021, 21, 7545 5 of 21

Sensors 2021, 21, x FOR PEER REVIEW

a Bangladeshi 5 of 21
native vehicle database with 10,440 images, divided into 13 categories.
Figure 2 presents the sample dataset images for each class.

Figure 1. The proposed method.

3.1. Deshi-BD Vehicle Dataset

The data source is an important component in supporting algorithms in learning fea-
tures and making predictions based on the learned information when using DL based
classification systems. To the best of our knowledge, there seems to be no standardized
Bangladeshi general vehicle dataset that includes data on Bangladeshi native vehicles and
that solves classification problems. As documented, CompCars as well as the Stanford
vehicle database simply incorporate the types of modern vehicles in specific areas. These
cannot be used in actual-time classifiers in other geographies, such as Bangladesh. In ad-
dition, the recommended data source differs from the available data sources. There is a
lot of variation regarding characteristics and structure. Furthermore, the present vehicle
classifier may not execute well enough in practical ITS implementations, as it is developed
on fairly short data samples with few classes [38]. To address these issues, we created a
Bangladeshi native vehicle database with 10,440 images, divided into 13 categories. Figure
Figure 1.
1. The proposed
proposed method.
presentsThe
2Figure method.images for each class.
the sample dataset
3.1. Deshi-BD Vehicle Dataset
The data source is an important component in supporting algorithms in learning fea-
tures and making predictions based on the learned information when using DL based
classification systems. To the best of our knowledge, there seems to be no standardized
Bangladeshi general vehicle dataset that includes data on Bangladeshi native vehicles and
that solves classification problems. As documented, CompCars as well as the Stanford
vehicle database simply incorporate the types of modern vehicles in specific areas. These
cannot be used in actual-time classifiers in other geographies, such as Bangladesh. In ad-
dition, the recommended data source differs from the available data sources. There is a
lot of variation regarding characteristics and structure. Furthermore, the present vehicle
classifier may not execute well enough in practical ITS implementations, as it is developed
on fairly short data samples with few classes [38]. To address these issues, we created a
Bangladeshi native vehicle database with 10,440 images, divided into 13 categories. Figure
2 presents the sample dataset images for each class.

SampleDeshi-BD
Figure2.2.Sample
Figure Deshi-BDdataset
datasetimages
imagesrepresenting
representingeach
eachclass.
class.

For this dataset, road surveillance and driving videos were gathered from Bangladeshi
transportations and highway roads in various weather conditions, such as daylight, foggy
day, and rainy day, and different lighting conditions, such as sunny, low light, and dark (in
the night) conditions, to properly extract the required images. In this study, thirteen native
Bangladeshi vehicle types were determined. The sizes of the objects in the collected images
range from very large objects, such as trucks, to little objects, such as traffic plants that are
difficult to detect with high accuracy [39]. After collecting the images, the database was
created by hand labeling with the help of a Windows snipping editing application. The
goal of creating this particular dataset is to build and analyze the collected Bangladeshi
traffic images. It will be used to increase our systems acceptability to enhance ITS in
South Asian nations, such as Bangladesh. The dataset contains 10,440 pictures, which are
classified into thirteen categories (auto rickshaw, bicycle, bus, car, CNG, cover van, easy
Figure 2. Sample Deshi-BD dataset images representing each class.
Sensors 2021, 21, 7545 6 of 21

bike, leguna, motorcycle, pickup, rickshaw, truck, and van). Figure 2 presents some sample
images in our Deshi-BD dataset and Table 1 shows image detail information of the dataset.
Data augmentation was applied on a low number of image classes to make our dataset
more robust.

Table 1. Data description of Deshi-BD vehicle dataset.

No. Vehicle Classes Data Augmentation Total Image

1 Auto Rickshaw 40 800
2 Bicycle 15 805
3 Bus 0 800
4 Car 20 865
5 CNG 15 830
6 Covered Van 30 810
7 Easy Bike 50 810
8 Leguna 0 760
9 Motorcycle 80 870
10 Pickup 20 740
11 Rickshaw 90 890
12 Truck 20 720
13 Van 0 740
Total 380 10,440

3.2. Data Preprocessing

Data preprocessing is critical to complete a deep learning-based classifier, such as
vehicle classification. This is because the vehicle images are collected from various sources
and, thus, data preprocessing is done to remove noise or unwanted background, to resize
the standard format image, and to make the vehicle images having uneven lighting system.
The preprocessing stage is divided into three separate parts:
i. Noisy/mislabeled vehicle image elimination;
ii. Vehicle images resizing;
iii. Augmentation.

3.2.1. Noisy/Mislabeled Vehicle Image Elimination

This section is explored unwanted background, noisy/mislabeled vehicles that will re-
duce the accuracy of prediction. First, we manually removed the noisy/mislabeled vehicle
images from our Deshi-BD vehicle dataset. Images from our Deshi-BD vehicle dataset were
categorized as (i) positive and (ii) negative for each class. In this way, Bangladeshi native
vehicle pictures can be ensured. Additionally, we could ensure the model’s efficiency, clas-
sified as positive and negative. Because the datum source was limited, data augmentation
methods were applied to increase test pictures and, therefore, enhance productivity of DL
architecture while avoiding overfitting issues.

3.2.2. Vehicle Images Resizing

Because data in the ImageNet dataset differ in terms of size, we created a standard size
for all images input into our DL model, and decreased resolution to “save” the preparation
and simulation period, as these systems must be evaluated for video classification.

3.2.3. Data Augmentation

When dealing with classification models, such as vehicle classification based on CNN
and DL architecture, it is essential to process image data. To address the issue of the
limited training data size, data augmentation was used [40]. This technique executes some
manipulations on the whole dataset. The goal was to create a collection of varied scenes,
therefore expanding the data. The DL method accomplishes perfect results in case of larger
datasets. Data augmentation is also used on drone image datasets to improve the accuracy
 3.2.3. Data Augmentation
When dealing with classification models, such as vehicle classification based on CNN
and DL architecture, it is essential to process image data. To address the issue of the lim-
ited training data size, data augmentation was used [40]. This technique executes some
Sensors 2021, 21, 7545 7 of 21
manipulations on the whole dataset. The goal was to create a collection of varied scenes,
therefore expanding the data. The DL method accomplishes perfect results in case of
larger datasets. Data augmentation is also used on drone image datasets to improve the
ofaccuracy of object identification
object identification and ensemble and models
ensemble models
[39]. [39]. We
We applied applied augmentation
augmentation to increaseto
increase
the the total
total images images
in our in our
dataset. This dataset.
resulted This resulted inthe
in permitting permitting
model to the model
train to train
successfully.
successfully.
Data augmentationData augmentation
is a strategy for is amaking
strategythe for entire
makingdatabase
the entiremore
database
robust.more
Asrobust.
a result,As
a result,
by by extending
extending the dataset,the dataset,
the methodthe method
decreases decreases overfitting
overfitting and improves
and improves generaliza-
generalization
tion ability.
ability. Here, Here,
what what is most
is most seriousisisthe
serious the second
secondissue.
issue.Data
Dataaugmentation
augmentation solves the prob-
solves the
lem without
problem causing
without any changes
causing to the to
any changes model’s structure.
the model’s However,
structure. Bangladeshi
However, native ve-
Bangladeshi
native vehiclecollections
hicle picture picture collections
are few. Theyare few.
are They are also challenging
also challenging to obtain to obtain
during theduring
COVID-19 the
COVID-19
pandemic.pandemic.
By
Byartificially
artificiallyboosting
boostingthe thesample
sampleusing usinglabel-preserving
label-preservingmodification
modificationalgorithms,
algorithms,
parametric
parametricdata dataaugmentation
augmentationisisthe thesimplest
simplestand andmost
mostfrequent
frequentway waytotoovercome
overcomethe the
problem of model overfitting [41]. We used a multiple augmentation
problem of model overfitting [41]. We used a multiple augmentation approach with the approach with the
vehicle
vehicleimages
images to to improve the the variation
variationof ofour
oursample:
sample:(i)(i) rotation,
rotation, (ii)(ii) horizontal
horizontal flip,flip,
(iii)
(iii) shifting
shifting (width
(width shift
shift andand height
height shift),
shift), (iv)(iv) zooming,
zooming, (v) (v) brightness
brightness adjustment,
adjustment, andand (vi)
(vi) shearing.
shearing. Figure
Figure 3 showsa adiagram
3 shows diagramofofthesethese improved
improved views.
views. The
The data
dataaugmentation
augmentation
phase
phaseaidsaidsininthe
thedevelopment
development ofof
a robust
a robust native vehicle
native classifier
vehicle utilizing
classifier utilizingminimal training
minimal train-
information and improves the efficiency of the DL algorithm. These
ing information and improves the efficiency of the DL algorithm. These augmentation augmentation methods
are relatedare
methods to the real to
related lifethe
scenario.
real life scenario.

Figure 3. Techniques of data augmentation performed on sample images.

Figure 3. Techniques of data augmentation performed on sample images.

3.3.Convolutional
3.3. ConvolutionalNeural
NeuralNetwork
Network(CNN)
(CNN)Model
Model
CNNisisaawell-known
CNN well-knowncutting-edge
cutting-edgeneural
neuralnetwork
networktechnology
technologythat thatisisuseful
usefulininCV
CV
tasks[42].
tasks [42]. CNN
CNNisisa atype
typeofof
deep neural
deep network
neural thatthat
network filters inputs
filters for meaningful
inputs infor-
for meaningful
mation using
information convolutional
using layers.
convolutional CNN’s
layers. CNN’sconvolutional
convolutionallayers apply
layers applyconvolutional
convolutional fil-
filters
ters totothe
theinput
input to
to compute
compute thethe output of neurons
neurons connected
connected to to particular
particularareas
areasininthe
the
input.
input.ItItis is
useful for for
useful extracting spatial
extracting and temporal
spatial characteristics
and temporal from images.
characteristics In CNN’sIn
from images.
convolutional layers, a weight-sharing mechanism is implemented to reduce the total
number of parameters [43,44]. CNN is usually made up of three layers: (i) a convolutional
layer for learning spatial and temporal features; (ii) pooling (a subsampling) layer for
reducing or subsampling the dimensionality of an input image; and (iii) a fully connected
(FC) layer for classifying the input image into various classes.

3.3.1. Transfer Learning

Transfer learning has been used successfully in a variety of applications, including
vehicle image classification and segmentation, in recent years. Transfer learning allows us
to learn a generic classifier, using a large amount of labeled data from the source domain
and a small amount of labeled data from the destination domain in classification problems.
In general, CNN performs better in large datasets than in smaller ones. When it is not
possible to create a large training dataset, transfer learning can be used. Figure 4 shows the
concept of transfer learning, where a model pre-trained on large benchmark datasets may
 Transfer learning has been used successfully in a variety of applications, including
vehicle image classification and segmentation, in recent years. Transfer learning allows us
to learn a generic classifier, using a large amount of labeled data from the source domain
and a small amount of labeled data from the destination domain in classification prob-
Sensors 2021, 21, 7545 lems. In general, CNN performs better in large datasets than in smaller ones. When8 of it 21
is
not possible to create a large training dataset, transfer learning can be used. Figure 4 shows
the concept of transfer learning, where a model pre-trained on large benchmark datasets
may be utilized
be utilized as a feature
as a feature extractor
extractor forrole,
for a new a new role,
using using a relatively
a relatively customsuch
custom dataset, dataset,
as a
such as a Deshi-BD
Deshi-BD dataset. dataset.

Figure 4.
Figure 4. Transfer
Transfer learning
learning concept.
concept.

In this research, transfer learning is used used toto solve

solve thethe challenge
challenge of of determining
determining the the
classes of Bangladeshi native vehicle vehicle images.
images. Because of the size and complexity complexity of of CNN
CNN
architectures, developing
developing and and testing
testing models
models could
could bebe costly
costly andand time-consuming.
time-consuming. When When
addressing
addressing aa vehicle
vehicle type
type classification,
classification, aa technique
technique called
called transfer
transfer learning
learning can
can give
give faster
faster
and more efficient
efficient outcomes.
outcomes. In transfer
transfer learning,
learning, weights,
weights, andand convolutional
convolutional filters
filters that
that
are capable of one task (learned for any classification
classification job) could be reused reused for another task
that requires just a little bit bit of
of retraining
retraining and and can
can be be learned
learned or or evaluated
evaluated on on aa limited
limited
number of images. Using a pre-trained neural network model with pre-loaded weights,
adjusting it to some amount, and then retraining part or the entire model to fulfil the new
task
task are
are examples
examples of of this.
this. The
The filters
filters trained
trained by by one
one task
task areare used
used toto extract
extract features
features from
from
images, which are
images, which are then
theninterpreted
interpretedby bythe theretrained
retrainedcomponent
componentofof thethe neural
neural network
network in
in order for it to complete its new task. In this research, the deep
order for it to complete its new task. In this research, the deep convolutional neural net- convolutional neural
network,
work, knownknown as as ResNet
ResNet [20],
[20], is isused
usedtotoexamine
examinetransfer
transferlearning,
learning, utilizing
utilizing pre-training
pre-training
over
over the “ImageNet” dataset [45], and the weights are the same as in ResNet
the “ImageNet” dataset [45], and the weights are the same as in ResNet [20].
[20].
Transfer
Transferlearning
learningisisperformed
performed bybyreplacing
replacingthethe
final fewfew
final layers of the
layers oforiginal network,
the original net-
including the output layer, with new fully connected layers that
work, including the output layer, with new fully connected layers that are appropriate are appropriate for the
for
new challenge. There are two methods to use transfer learning
the new challenge. There are two methods to use transfer learning from a model: reuse from a model: reuse the
model as a as
the model feature extractor
a feature and apply
extractor a completely
and apply different
a completely classifier,
different or reuse
classifier, orthe model
reuse the
to do fine-tuning. Fine-tuning is an approach that uses unfrozen
model to do fine-tuning. Fine-tuning is an approach that uses unfrozen layers of a com- layers of a complete
model to slightly
plete model changechange
to slightly both the bothnew thefully
newconnected layers oflayers
fully connected the classifier and specific
of the classifier and
CNN layers, such as convolutional layers [46]. Transfer learning has started from fully
specific CNN layers, such as convolutional layers [46]. Transfer learning has started from
connected layer because of a fully connected (FC) layer for classifying the input image into
our target classes.

3.3.2. Convolutional Layer

This is the initial layer and one of the core parts of a convolutional neural network
(CNN). Definite sets of learnable filters are present in this layer, regarded as a supreme layer
in the CNNs. Spatially—input-sizes are greater than the filters. During the forward pass,
these filters slip across the input feature to produce a 2D feature map that indicates location
as well as the strength of the identified visual elements in data source. The following
formula is used to calculate the characteristics of these layers:
 
l l −1
yln = f l ∑ m −→ y (1)
n m
Sensors 2021, 21, 7545 9 of 21

l
Here, the nth feature map of l-layer is yln , C-kernel is m −→ , while feature
n
extraction from layer-l and ylm−1 is the characteristic pattern connected to layer-l.

3.3.3. Pooling Layer

Pooling layers use the image building connection concept to execute pooling pro-
cedures on feature maps, in order to reduce network congestion while holding the key
characteristics. They are commonly utilized among the CNN model consecutive convo-
lutional layers, and are performed to slowly reduce the spatial display space, decreasing
operations while preserving crucial data, which helps in minimizing overfitting during the
training process.
Average pooling: as the filter moves over the input, it calculates the average value
inside the receptive field to send to the output array.
Maximum pooling: as the filter passes over the input, it picks the pixel with the
highest value to transfer to the output array. As an aside, this technique is more commonly
used than average pooling. The pooling function can be implemented through:
 
yln = f l zln−1 xwnl + bnl (2)

Here, zln−1 value is extracted from l − 1 convolution features, wlx represents map
weight, and bxl represents offset value.

3.3.4. Dropout Layer

While training the CNN model, we observed a significant amount of overfitting. Thus,
to lessen the influence of overfitting, the dropout layer was deployed. Another popular
approach used in CNN, to sidestep consequences of overfitting, is regularization. This
is accomplished by applying a substantial charge to the loss function in use. Hence, the
particular layer is included in the final recommended system, which aids in preventing the
system from becoming dependent on the other feature weights.

3.3.5. Fully Connected Layer

The conventional method for image classification problems is to utilize a stack of fully-
connected layers, followed by a softmax activated layer [37]. The probability distribution
over each possible class label is generated by the softmax layer, and then we simply classify
the image based on the most possible class. The last part of the CNN architecture is the fully
connected layer, after the two stages of alternate convolution, batch normalization, ReLU,
and pooling sublayers. In order to minimalize the feature dimensions, neurons inside this
layer are linked to every activation in the former layer. These layers are essentially just
like a normal neural network. Here, they map the flattened data into the class labels and
generate values for each output variable. Eventually, overall the results of these layers are
supplied as inputs for a softmax layer, where the values are transformed, ranging between
0 to 1 and ending in a total of 1. In this approach, the softmax layer depicts the result as a
real probability. Fully connected neurons can be described mathematically as follows:
 
N
yln = f l ∑ml=−11 ylm−1 wm,n
l
+ bnl (3)

Here, Nl represents value of neurons of output layer, ylm−1 represents m characteristic

l
pattern of layer l − 1, and wm,n represents connected weights.

3.4. AlexNet
For the first time, AlexNet was introduced in 2012, which used an eight-layer CNN
model. This model won the ImageNet Large Scale Visual Recognition Competition by
an extraordinarily great margin. AlexNet demonstrated how learning-derived features
might outperform manually generated features, shattering the prior CV standard. Al-
Sensors 2021, 21, 7545 10 of 21

though AlexNet [36] is a commonly used deep CNN network, it may still achieve viable
classification efficiency when compared to other types of networks. During this model’s
training step, the input data are scaled to 224 × 224 pixels and fed into the system. The
AlexNet design initially utilizes a convolutional layer to conduct convolutional as well
as max pooling through local response normalization utilizing 96 distinct size 11 × 11
receptive filters. Max-pooling activities were carried out using 3 × 3 filters with a stride
size of 2. Similar processes were carried out in the second layer with 5 × 5 filters. Moreover,
3 × 3 filters were utilized, as well as fourth and fifth convolutional layers with 384, 384,
and 296 feature maps, correspondingly. The output of the two fully linked layers is utilized
as a feature extraction vector with dropout, then a softmax layer at the finish point.

3.5. VGGNet
VGG-16 is a CNN model [37] with thirteen convolutional layers and three fully-
connected layers for a total of 16 weight layers. More exactly, the size of VGG-16 trained
ImageNet weights is 528 MB, and VGG16 contains 138 million parameters in total. There-
fore, it takes quite a lot of disk space and bandwidth that makes it inefficient. In contrast to
AlexNet, this VGGNet architecture contains numerous parameters. Moreover, VGGNet re-
quires a lot of memory, which makes it more expensive computationally. In spite of being a
deep network, and having huge complexity in computing, the model outperforms AlexNet
and GoogLeNet in terms of productivity. Moreover, it is very simple to put into action. The
ILSVRC-2014 challenge on ImageNet, for 1000 classes, scored a 92.70% precision rate.

3.6. ResNet
The residual network (ResNet) is one of the most widely utilized CV architectures [20].
There are many benefits of using ResNet; the prominent benefit is that it can resolve the
difficulty of degrading accuracy and the vanishing gradient by familiarizing the idea of
shorter links. As a result, it is adaptable, task-specific, and capable of preparing very
deep learning algorithms. Residual nets on the ImageNet dataset [47] have a depth of
about 152 layers; these are eight times the depth of VGG networks. These, however, have a
decreased risk of complications. On the ImageNet test set, a collection of these residual
nets achieves a 3.57% error rate. On the tasks of ImageNet detection, ImageNet localization,
COCO detection, and COCO segmentation (“The COCO dataset is a large-scale object
detection, segmentation, and labeling dataset published by Microsoft. COCO has several
features. Object segmentation, recognition in context, superpixel stuff segmentation, 330 K
images (>200 K annotated), 1.5 million object instances, 80 object categories, 91 stuff cat-
egories, 5 captions per image, 250,000 persons with coordinates. Machine Learning and
computer vision developers frequently utilize the COCO dataset for various computer
vision applications”), deep residual networks came in first. ResNet-50 features 48 convo-
lutional layers, as well as one max pooling and one average-pooling layer. The overall
number of weighted layers is 50, with a total of 25, 583, 592 trainable parameters. It can
Sensors 2021, 21, x FOR PEER REVIEW 11 of 21
do 3.8 × 109 floating-point computations. Figure 5 depicts the architecture of the original
ResNet-50.

Figure5.5.AApre-trained
Figure pre-trainedResNet-50
ResNet-50network
networkwith
withresidual
residualconnections.
connections.

We
Weutilized
utilizedResNet-50
ResNet-50as
asthe
thebasic
basicmethodology
methodologyininouroursuggested
suggestedmodel,
model,which
whichwaswas
pre-trained
pre-trainedon on the
the ImageNet dataset [45]
[45] for
forimage
imageclassification.
classification.We
Wemoved
movedthe
the initial
initial 49
49 layers of ResNet-50, which were kept frozen on the classification model, adopting
layers of ResNet-50, which were kept frozen on the classification model, adopting transfer
learning techniques [48]. All of the other layers may be categorized as learnt feature ex-
traction layers, which produce bottleneck features as the activation maps. We train a 13-
fully connected softmax using the bottleneck characteristics of our native vehicle images
as inputs, since we have 13 classes, and then swap the 1000 fully connected softmax with
 Figure 5. A pre-trained ResNet-50 network with residual connections.

Sensors 2021, 21, 7545 11 of 21

We utilized ResNet-50 as the basic methodology in our suggested model, which was
pre-trained on the ImageNet dataset [45] for image classification. We moved the initial 49
layers of ResNet-50, which were kept frozen on the classification model, adopting transfer
learninglearning
transfer techniques [48]. All [48].
techniques of theAll
other layers
of the may
other be categorized
layers as learnt feature
may be categorized ex-
as learnt
tractionextraction
feature layers, which produce
layers, whichbottleneck features asfeatures
produce bottleneck the activation maps. Wemaps.
as the activation train aWe
13-
fullyaconnected
train softmax using
13-fully connected softmaxtheusing
bottleneck characteristics
the bottleneck of our native
characteristics of our vehicle images
native vehicle
as inputs,
images since we
as inputs, have
since we13 classes,
have and then
13 classes, andswap the 1000
then swap the fully connected
1000 fully softmax
connected with
softmax
our trained
with data,data,
our trained which can be
which canseen in Figure
be seen 5. 5.
in Figure

3.7.
3.7.Proposed
ProposedClassification
ClassificationModel
Model
We
Wesuggested
suggestedaapipeline-based
pipeline-basedtechnique
techniqueforfor
our
ourBangladeshi
Bangladeshi native
nativevehicle classifica-
vehicle classifi-
tion model for our DL model. The pipeline was divided into multiple
cation model for our DL model. The pipeline was divided into multiple phases, the phases, the firstfirst
of
which received images of Bangladeshi native vehicles from our Deshi-BD
of which received images of Bangladeshi native vehicles from our Deshi-BD vehicle da- vehicle dataset,
and theand
taset, lastthe
of last
which classified
of which the model.
classified The output
the model. of eachofstep
The output eachwas used
step wasasused
the input
as the
for
input for the next stage. The suggested training method is divided into three (i)
the next stage. The suggested training method is divided into three stages: image
stages: (i)
acquisition and processing; (ii) model selection and training; and (iii) evaluation.
image acquisition and processing; (ii) model selection and training; and (iii) evaluation. Figure 6
represents the suggested pipeline technique.
Figure 6 represents the suggested pipeline technique.

Figure6.6.Proposed
Figure Proposedmodel
modelapproach
approachpipeline.
pipeline.

3.7.1.Image
3.7.1. ImageAcquisition
Acquisitionand
andPreprocessing
Preprocessing
Tocreate
To createour
ourDeshi-BD
Deshi-BDvehicle
vehiclecollection,
collection,we wegathered
gatheredimages
imagesofofBangladeshi
Bangladeshinative
native
vehicles.
vehicles.The
Thedataset
datasetrequired
requiredpreprocessing,
preprocessing,image imagescaling,
scaling,andandappropriate
appropriateconversions
conversions
for
forthe
theDLDLmodel.
model. The The amount
amount ofof images
images in in our
our sample
sample werewere not
notthe
thesame
samesize
sizefor
forall
all
classes,
classes,which
whichcaused
causedour ourdataset
datasettotobeberelatively
relativelyunbalanced.
unbalanced. To Tosidestep
sidestepthis
thisproblem,
problem,
we
weused
usedseveral
severalaugmentation
augmentationapproaches
approachesto toincrease
increasethe theamount
amountof ofimages
imagesin inour
ourmodel,
model,
which
which allowed it to achieve greater generalization and recognition. After loadingimages
allowed it to achieve greater generalization and recognition. After loading images
into
intoour
ourchosen
chosenmodel
modelfromfromthe
theDeshi-BD
Deshi-BDvehicle
vehicledataset,
dataset,thetheimages
imageswerewerefirst
firstdivided
divided
into
intotraining
trainingand
andvalidation
validationdata,
data,according
according to to
thethe
CNNCNN architecture’s typical
architecture’s input
typical sizesize
input of
normalized
of normalizeddatadata
to the sizesize
to the of 224 × 224
of 224 pixels.
× 224 pixels.

3.7.2. Model Selection and Training

A DL method takes, as input data, a map of characteristics to the target “X”, and
predicts a model based on the output “Y”. ResNet-50 architecture was used for our model.
During training, the algorithm optimized the parameters (update weights and biases) that
were utilized for our model’s identification. In our experiment, we used 80% image data
for our training model. Moreover, for validating the model, 20% image data were allocated
to form a validation subset. On the test set, the performance of the suggested deep learning
model was evaluated.

3.7.3. Evaluation
Following the completion of our model’s training, the proposed DL model’s per-
formance was assessed using a variety of assessment measures, including (i) accuracy;
(ii) precision; (iii) recall; and (iv) F1 − Score.

3.8. Confusion Matrix as Evaluation Metrics

The confusion matrix is gathered in order to illustrate the predictions made by the
developed framework on the testing data and to identify the bunch of frames incorrectly
Sensors 2021, 21, 7545 12 of 21

categorized. We investigated the correct execution of the native vehicle classification model
in relation to model classifier indexes. The efficacy of our proposed native recognition and
classification system is evaluated by generating evaluation metrics based on four major
impacts used to test the classifier: true positives (Tp ), true negatives (Tn ), false positives
(Fp ), and false negatives (Fn ). The overall acceptability of the fraction of the native vehicle
classification model that is properly categorized is shown below:
Model accuracy (Acc ) improves the capacity to identify the Bangladeshi native vehicle
categories properly. To assess the correctness of a testing dataset, we compute the percent-
age of true positive (Tp ) and true negative (Tn ) instances determined by given contacts:

Tp + Tn
Acc = (4)
Tp + Fp + Tn + Fn

Here, true positive ( Tp ) shows the number of expected positive classes that are really
positive classes, true negative ( Tn ) counts the set of anticipated negative classes that are
really negative classes, false positive ( Fp ) displays the quantity of genuine negative classes
anticipated as positive classes, false negative ( Fn ) displays the quantity of real positive
classes anticipated as negative classes. Additionally, model accuracy Acc is a ratio of
appropriately predicted observations to all inspections. It is excellent when the datasets
are symmetric.
Model precision ( Prc ) is the ratio of appropriately predicted observations of the total
predicted positive assessment. Model precision ( Prc ) symbolizes the total amount of
real positive class images among all classes images anticipated to be positive. It may be
calculated as:
Tp
Prc = (5)
Tp + Fp
Model recall ( Rec ) is ratio of appropriately predicted clarifications to all assessments
in the real class. The recall signifies all positive class image ratios, which are magnificently
projected as positive. It may be calculated as:

Tp
Rec = (6)
Tp + Fn

F1 − Score is a harmonic mean of Prc and Rec , and so it provides a combined knowledge
about these two metrics. When the class distribution is imbalanced, F1 − Score is more
beneficial than accuracy. The F1 − Score is maximum when Prc and Rec are equal.

2 × Prc × Rec
F1 − Score = (7)
Prc + Rec

4. Result
4.1. Experiment Setup
The dataset-based platform configuration is used to test our proposed Bangladeshi
native vehicle classification model. Our suggested tests were carried out on a heavy com-
puting machine with an NVIDIA GeForce RTX2060 GPU, a DDR5 8 GB graphics processing
unit (GPU), and a 64-bit Windows 10 operating system, with an Intel Core i7-8750H @
2.2 GHz CPU and 16 GB RAM. We utilized Python 3.6 to develop our classification model,
with the Keras Library being frontend and TensorFlow being backend for the framework.

4.2. Experimental Outline

We used the ResNet-50 model in the experimental framework to conduct Bangladeshi
native vehicle classification. CNN architecture is one of the most innovative plans devel-
oped by He et al. [20], and it took top position in the ILSVRC-15 with a best-five inaccuracy
ratio of 3.57% by demonstrating amazing results in object recognition and classification [34].
In early DL networks, growing network layers might cause a vanishing gradient issue,
 brary being frontend and TensorFlow being backend for the framework.

4.2. Experimental Outline

We used the ResNet-50 model in the experimental framework to conduct Bangla-
Sensors 2021, 21, 7545 deshi native vehicle classification. CNN architecture is one of the most innovative plans 13 of 21
developed by He et al. [20], and it took top position in the ILSVRC-15 with a best-five
inaccuracy ratio of 3.57% by demonstrating amazing results in object recognition and clas-
sification [34]. In early DL networks, growing network layers might cause a vanishing
preventing the model
gradient issue, fromthe
preventing converging
model from at itsconverging
optimum. atInits theoptimum.
ResNet model,
In the aResNet
unique
skip connection-based method was presented, in which every input
model, a unique skip connection-based method was presented, in which every input from from the last layer
was collected and delivered to the result of the following layer. Meanwhile,
the last layer was collected and delivered to the result of the following layer. Meanwhile, to reduce time
complexity, a bottleneck suggestion was incorporated in the deeper network,
to reduce time complexity, a bottleneck suggestion was incorporated in the deeper net- driving this
CNN
work, driving this CNN model. We have been experimenting with a transfer learning ap- a
model. We have been experimenting with a transfer learning approach in which
model
proachtrained
in whichforaamodel
fairly precise
trainedassignment may be assignment
for a fairly precise modified tomay implement an alternative
be modified to im-
assignment
plement an by simply learning
alternative assignmentthe new weights.
by simply In spite
learning theofnew
having a smaller
weights. dataset,
In spite which
of having
isa not sufficient
smaller forwhich
dataset, training from
is not the beginning,
sufficient this from
for training technique is highly this
the beginning, effective. Figure
technique is 7
shows the overall framework of our customized DL model.
highly effective. Figure 7 shows the overall framework of our customized DL model.

Figure 7.
Figure 7. An
An overview
overview of
of the
the proposed
proposed model
modeloutline.
outline.

The Deshi-BD vehicle dataset was used to train ResNet-50 pre-trained architecture.
A dropout layer is present at the bottom of the model to solve the vanishing gradient
problem. A new, fully connected layer was added based on the classification block to
conduct thirteen categories of native vehicle classifications, where each unit in the last layer
was linked to the thirteen-class output probability using the softmax function. To confirm
that these newly added layers learn greater-level feature characteristics from the dataset,
we enlarged their learning rate in comparison to the prior layers’ learning rate. In addition,
only the specific newly added layers included in the base model were trained, with the
early convolutional layers frozen. The main concept behind freezing these layers was
to accelerate convergence while avoiding the gradient outbreak throughout the training
phase. After removing texture characteristics, classification was performed to compare the
projected class to the real class. Over the training procedure, the network’s computation
costs decreased, while the total trainable parameters of the modified CNN model also
reduced. The architecture of our proposed ResNet-50 pre-trained CNN model is shown
in Table 2.
Sensors 2021, 21, 7545 14 of 21

Table 2. Architecture of our proposed ResNet-50 pre-trained CNN model.

Layer Name Output Size Layers

Conv1 7 × 7, 64, stride 2
112 × 112
Pooling 3 × 3maxpool, stride 2
1 × 1, 64
 
Conv2 56 × 56  3 × 3, 64  × 3
1 × 1, 256
1 × 1, 128
 
Conv3 28 × 28  3 × 3, 128  × 4
1 × 1, 512
1 × 1, 128
 

Conv4 14 × 14  3 × 3, 128  × 6
1 × 1, 512
1 × 1, 128
 

Conv5 7×7  3 × 3, 128 ×3

1 × 1, 512
Adaptive average pooling 2D
fc1 : Input feature = 1024, Output feature = 512
ReLU (in place), dropout = 0.5
Pooling
fc2 : Input feature = 512, output feature = 13
Our implemented 1×1
Softmax( )
classification block
Classification output = (loss
=‘categorical_crossentropy’,
optimizer = ‘adam’, metrics = (‘accuracy’))

In our experiment, we used a transfer learning approach on a pre-trained ResNet-50

architecture to recycle the weights of the network learnt from ImageNet, as shown in
Figure 7. To avoid poor initialization, pre-trained weights were used in the model, as
with its counterpart “random initialization of weights”. This network has 50 layers of
depth, which was gained by substituting each 2-layer block in the innovative ResNet with
the 3-layer bottleneck block [20]. This network’s input layer accepted red–green–blue
(RGB) color images, reshaped to 224 × 224 pixels. To implement transfer learning, the
network’s last fully connected layer, which performed ImageNet classification, was deleted.
The pre-trained model’s early convolutional layers served as a foundation network for
the newly adapted model. Following a universal mean pooling layer, two sets of batch
normalization: (i) fully connected, and (ii) dropout layers, were added to the base network.
The first fully connected layer has 512 neurons and the other has 256 neurons. Each
fully connected layer was trailed by a ReLU activation layer. The training process of the
pre-trained model is reduced by adding the batch normalization layers. The problem
of overfitting was inherently reduced by the inclusion of global average pooling. The
problem of overfitting in deep models typically fails to have a decent generalization on
input that has not been seen earlier (test data). We employed numerous data augmentation
techniques to prevent overfitting in our dataset. Finally, the final layer of our proposed
model classified the Bangladeshi native vehicle images into thirteen classes using the
softmax activation function.
TensorFlow resources were used to load the AlexNet, VGG-16, and ResNet-50 models
for evaluation. TensorFlow was used to train all networks. In our proposed native vehicle
classification model, the parameters are optimized using a stochastic optimization approach,
called the Adam optimizer. We used dropout ratios of 0.50 for both dropout layers and set
our learning rate at 0.0001. To train the model, we used a batch size of 32 and 50 epochs.
Categorical cross-entropy, a commonly used loss function, was used to accumulate loss
during the process, and validation of the network was performed after every epoch to
evaluate the learning. Despite the fact that our proposed model was trained for 50 epochs,
it obtained validation accuracy of 92% for AlexNet, 95% for the VGG-16 pre-trained model,
 els
els for
for evaluation.
evaluation. TensorFlow
TensorFlow was was used
used to
to train
train all
all networks.
networks. In In our
our proposed
proposed native
native ve-
ve-
hicle
hicle classification
classification model,
model, the
the parameters
parameters are
are optimized
optimized using
using aa stochastic
stochastic optimization
optimization
approach,
approach, called
called the
the Adam
Adam optimizer.
optimizer. We
We used
used dropout
dropout ratios
ratios of
of 0.50
0.50 for
for both
both dropout
dropout
layers
layers and set our learning rate at 0.0001. To train the model, we used a batch size
and set our learning rate at 0.0001. To train the model, we used a batch size of
of 32
32
and 50 epochs. Categorical cross-entropy, a commonly used loss
and 50 epochs. Categorical cross-entropy, a commonly used loss function, was used function, was used to
to
Sensors 2021, 21, 7545 15 of 21
accumulate loss during the process, and validation of the network was
accumulate loss during the process, and validation of the network was performed after performed after
every
every epoch
epoch toto evaluate
evaluate the
the learning.
learning. Despite
Despite the
the fact
fact that
that our
our proposed
proposed model
model waswas trained
trained
for
for 50 epochs, it obtained validation accuracy of 92% for AlexNet, 95% for the VGG-16
50 epochs, it obtained validation accuracy of 92% for AlexNet, 95% for the VGG-16
and 98% for the ResNet-50
pre-trained pre-trained model. Figure 8 displays the relative accuracy theof
pre-trained model,
model, and and 98%
98% for
for the
the ResNet-50
ResNet-50 pre-trained
pre-trained model.
model. Figure
Figure 88 displays
displays the
various
relative networks.
accuracy of various networks.
relative accuracy of various networks.

98%
100% 95% 98%
100% 92% 95%
Accuracy
92%
Accuracy
95%
95%
90%
90%
85%
85% AlexNet VGG-16 ResNet-50
AlexNet VGG-16 ResNet-50
Accuracy 92% 95% 98%
Accuracy 92% 95% 98%
CNN model
CNN model

Figure8.8.Validation
Validation accuracyof
of selectedCNNs
CNNs model.
Figure 8. Validationaccuracy
Figure accuracy ofselected
selected CNNsmodel.
model.

In
In addition,
Inaddition, for
foraaabetter
addition,for better understanding
betterunderstanding
understandingof of our
ofour approach,
ourapproach, the
approach,the learning
thelearning process
learningprocess
processfor for both
forboth
both
training
training and validation is illustrated using a loss and accuracy curve. We draw the loss
training and
and validation
validation is
is illustrated
illustrated using
using aa loss
loss and
and accuracy
accuracy curve.
curve. We
We draw
draw the
the loss
loss
and
and accuracy
andaccuracy curves
accuracycurves
curvesforfor each
each
for each classification
classification andand
classification recognition
recognition
and network
network
recognition model
model
network throughout
throughout
model throughoutthe
training
the phase. Figure 9 illustrates the model’s loss andandaccuracy values as a function of
the training
training phase.
phase. Figure
Figure 99 illustrates
illustrates the
the model’s
model’s loss
loss and accuracy
accuracy values
values as
as aa function
function
training
of epochs.
training epochs.
of training epochs.

Figure 9. Accuracy and loss curve for training and validation.

Figure 9. Accuracy and loss curve for training and validation.
Figure 9. Accuracy and loss curve for training and validation.

As seen above, the ResNet-50 model outperformed AlexNet and VGG-16 in terms
of accuracy, with a 3% difference on average. As a result, after fine-tuning the design,
ResNet-50 is expected to attain more accuracy.

4.3. Performance Analysis

The confusion matrix is an excellent assessment directory for classification prob-
lems [49]. In a confusion matrix, rows and columns represent the real and projected classes.
 As seen above, the ResNet-50 model outperformed AlexNet and VGG-16 in terms of
accuracy, with a 3% difference on average. As a result, after fine-tuning the design, Res-
Net-50 is expected to attain more accuracy.

Sensors 2021, 21, 7545 4.3. Performance Analysis 16 of 21

The confusion matrix is an excellent assessment directory for classification problems

[49]. In a confusion matrix, rows and columns represent the real and projected classes. The
recognition
The accuracy
recognition is represented
accuracy in this
is represented matrix
in this of prediction
matrix results
of prediction by a by
results blue colorcolor
a blue box,
and the deeper the color, the more accurate the model recognition [50]. The
box, and the deeper the color, the more accurate the model recognition [50]. The horizontal horizontal axis
axis represents the predicted values of the test set, such as auto rickshaw, bicycle, bus, car,
represents the predicted values of the test set, such as auto rickshaw, bicycle, bus, car,
CNG, cover
CNG, cover van,
van, easy
easy bike,
bike, leguna,
leguna, motorcycle,
motorcycle, pickup,
pickup, rickshaw,
rickshaw, truck,
truck, and
andvan.
van. The
The
vertical axis
vertical axis represents
represents thethe real
real values
values ofof the
the test
test samples,
samples, such
such asas auto
auto rickshaw,
rickshaw, bicycle,
bicycle,
bus, car, CNG, cover van, easy bike, leguna, motorcycle, pickup, rickshaw
bus, car, CNG, cover van, easy bike, leguna, motorcycle, pickup, rickshaw truck, and van. truck, and van.
The projected value of the model that is consistent with the true value
The projected value of the model that is consistent with the true value of the test sampleof the test sample
is located
located on
on the
the crosswise
crosswise axisaxis of
of the
the matrix.
matrix. Figure
Figure 1010 show
show thethe findings
findings ofof our
our model’s
model’s
confusion matrix.
confusion matrix.

Figure 10.
Figure 10. Confusion
Confusionmatrix
matrixfor
forDeshi-BD
Deshi-BDnative vehicle
native dataset;
vehicle (a) (a)
dataset; confusion matrix,
confusion without
matrix, nor-
without
malization, and (b) normalized confusion matrix.
normalization, and (b) normalized confusion matrix.

In this part, we utilize a variety of evaluation parameters to assess the model clas-
sification outcomes, including precision, recall, and F1 − Score results for our developed
framework on the testing data. Equations (4)–(6) are used to calculate the outcome (Model
Evaluation Metrics). Table 3 provides a thorough performance analysis of the classification
impact based on each category. As it can be observed, the classification result of our chosen
model had a 98% average precision, recall, and F1 − Score.
Sensors 2021, 21, 7545 17 of 21

Table 3. The classification report for our proposed ResNet-50 pre-trained model.

Vehicle Type Precision Recall F1 −Score Number of Test Images

Van 0.98 0.98 0.98 129
Rickshaw 0.98 0.98 0.98 182
Leguna 0.99 0.97 0.98 156
Bicycle 0.99 0.99 0.99 162
Truck 0.97 0.98 0.97 131
CNG 0.98 0.99 0.99 159
Pickup 0.96 0.95 0.95 133
Motorcycle 0.99 0.98 0.99 188
Car 0.98 1 0.99 175
Covered Van 0.98 0.98 0.98 174
Bus 0.98 0.98 0.98 162
Auto Rickshaw 0.96 1 0.98 161
Easy Bike 0.99 0.97 0.98 176

We saw that Figure 8 represents the comparative experiment findings in this study;
moreover, we compare the findings with two other CNN models; AlexNet and VGG-16
pre-trained architecture. There has been a significant improvement in model performances
in terms of accuracy and other assessment criteria. As shown in Table 4, our presented
model outperformed the AlexNet and pre-trained VGG-16 models.

Table 4. Report on comparative performance analysis.

Model Type Precision Recall F1 −Score Accuracy

AlexNet 0.915 0.935 0.935 0.928
VGG-18 0.942 0.945 0.951 0.946
ResNet-50 0.979 0.981 0.981 0.98

The AlexNet model only obtained a classification accuracy of almost 93%, while the
VGG-16 model achieved a classification accuracy of nearly 95%, which was around 6%
and 3% lower than the accuracy attained by our suggested ResNet-50 pre-trained model,
respectively. Aside from that, the most probable reason seemed to be that our suggested
ResNet-50 pre-trained model had less trainable parameters than VGGNet (134 M), and
VGGNet did not include skip connections to make calculations easier. Built on the find-
ings, we can conclude that our suggested model, which is based on the residual network
ResNet-50, is more robust for the Bangladeshi native vehicle classification-based CNN
model that can extract robustness for recognition and classification. As can be observed,
our proposed ResNet-50 pre-trained DL model, based on transfer learning, outperforms
the other three classification methods, in terms of classification accuracy.

5. Discussion
The proposed Bangladeshi native vehicle classifier was related to the existing vehi-
cle classifiers [16,51–53] to verify the selected network’s efficacy. The existing networks
have been reproduced on the suggested database. The GoogLeNet architecture-based
vehicle classification method and the 22-layer depth network, was presented by Zhuo
et al. [51]. Gao et al. [52] developed an AlexNet that focuses on vehicle classifier using
five convolutional layers as well as three fully connected layers. Additionally, an incep-
tion architecture-based classifier was suggested by Zakria et al. [16]. A self-proposed
CNN-based vehicle classifier with 13 convolutional layers, as well as one fully connected
layer, max-pooling, as well as dropout layers, is followed in the model, which was pre-
sented by Shivai et al. [53]. All of these models established good performances on their
datasets. However, these existing systems [16,51–53] did not focus on massive data be-
cause they are constructed of many smaller depth networks and are developed on narrow
Sensors 2021, 21, 7545 18 of 21

categories that do not include road transport mobility. Table 5 compares this study to other
baseline techniques.

Table 5. A comparison of the proposed approach to existing vehicle classifiers.

Authors Features Model Accuracy (%)

Zhuo et al. [51] GoogLeNet based GoogLeNet 95.49
Gao et al. [52] AlexNet based AlexNet 92.61
Shivai et al. [53] CNN based CNN 88.96
Zakria et al. [16] Inception based Inception 92.77
Our proposed model ResNet-50 based Pre-trained model 98

As a result, for the native vehicle classification problem, the above techniques perform
poorly in real-time classification purposes. Furthermore, it is crucial to note that these
methods are trained on imbalanced data, which is a key element in the true presentation of
local vehicle classification architecture. As a result, the achievement of the current models
is discriminating when assessing the presented Deshi-BD stable database. However, our
proposed native vehicle classification network was trained on the own-built Deshi-BD
native vehicle database, which covers thirteen common road traffic classes in Bangladesh,
with 10,440 images. As a result of evaluating all of the aforementioned systems, we
conclude that our suggested native vehicle classifier outperformed the other existing vehicle
classifiers, in terms of accuracy. Additionally it has greater generalization capabilities for
taking a broader variety of information, as well as adaptability to South Asian traffic
surveillance applications. In the future, we hope to expand our research and develop fine-
grained classification systems that will increase the usefulness of the proposed approach
in ITS.

6. Conclusions
Road traffic accidents are major causes of death and injury across the world, which
is a concerning problem. Thousands of individuals are killed or seriously disabled in
traffic crashes every year. When compared to other developing countries, Bangladesh has
a much higher percentage of death and injury from traffic accidents. To avoid this serious
problem, in this paper, we presented a customized DL model-based method to detect and
classify Bangladeshi native vehicles, to develop the effectiveness of ITS. In our work, for
the training classification algorithm, a new dataset was introduced for Bangladeshi native
vehicles, namely the Deshi-BD dataset, having 10,440 images with thirteen categories to
train the classification system. At first, to validate the performance of our Deshi-BD vehicle
dataset, three advanced CNN architectures, AlexNet, VGG, and ResNet, were trained. The
ResNet-50 model was constructed based on transfer learning for the Bangladeshi native
vehicle classification. Transfer learning was used to improve the ResNet-50 architecture
by adding a new classification layer to the original network. In our model, we also
evaluated the performance by using various attributes, such as (a) accuracy, (b) precision,
(c) recall, and (d) F1 − Score. We also compared our model with VGGNet and AlexNet.
Results from our study exposed that our recommended native vehicle classification system
achieved accuracy of 98% on the Deshi-BD dataset, which is significantly greater than other
existing advanced classification systems. This methodology proved to be “vigorous” while
deviations, such as vehicle size, vehicle shade, position, and weather, were considered.
Moreover, we proposed native vehicle classification methods that utilize fewer parameters,
so that training costs can be lower and, thus, have the prospective to support intelligent
traffic management technologies. Finally, we conclude that our approach outperforms
other current methods in vehicle type classification under all conditions.
However, in order to improve our research, we need to gather relevant data and
evaluate other vehicles that are also visible on the road. Our future work will involve
adding more types of native vehicle images to the dataset, as well as further adjusting
the network topology and parameters for vehicle classification. We plan to expand our
Sensors 2021, 21, 7545 19 of 21

model in real-time detection and tracking for smart traffic monitoring system. We also plan
to grow our system to include vehicle counting, automatic license plate recognition, and
traffic congestion detection modules before combining them into a full autonomous traffic
monitoring system.

Author Contributions: Conceptualization, M.M.H. and K.F.; methodology, M.M.H.; software,

M.A.I.H. and M.M.H.; validation, M.M.H., M.A.I.H. and K.F.; formal analysis, M.M.H. and M.A.I.H.;
investigation, Z.W. and M.A.I.H.; resources, M.M.H. and Z.W.; data curation, M.M.H. and K.F.;
writing—original draft preparation, M.M.H. and K.F.; writing—review and editing, Z.W. and
M.A.I.H.; visualization, M.M.H. and K.F.; supervision, Z.W. All authors have read and agreed
to the published version of the manuscript.
Funding: 2019YFC1521300: National Key Research and Development Program of China.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Acknowledgments: This work was supported by the National Key Research and Development
Program of China (2019YFC1521300). The authors would also like to thank the College of Information
Science and Technology, Donghua University, Shanghai, 201620, China, for providing us with
resources and support.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Singh, S. Road Traffic Accidents in India: Issues and Challenges. Transp. Res. Procedia 2017, 25, 4708–4719. [CrossRef]
2. Islam, M.A.; Dinar, Y. Evaluation and Spatial Analysis of Road Accidents in Bangladesh: An Emerging and Alarming Issue.
Transp. Dev. Econ. 2021, 7, 10. [CrossRef]
3. Alam, M.S.; Mahmud, S.S.; Hoque, M.S. Road accident trends in Bangladesh: A comprehensive study. In Proceedings of the 4th
Annual Paper Meet and 1st Civil Engineering Congress, Dhaka, Bangladesh, 22–24 December 2011.
4. Shafabakhsh, G.A.; Famili, A.; Bahadori, M.S. GIS-based spatial analysis of urban traffic accidents: Case study in Mashhad, Iran.
J. Traffic Transp. Eng. 2017, 4, 290–299. [CrossRef]
5. Szeliski, R. Computer Vision: Algorithms and Applications, 1st ed.; Springer Science & Business Media: Berlin/Heidelberg,
Germany, 2010.
6. AbdelBaki, H.M.; Hussain, K.; Gelenbe, E. A laser intensity image based automatic vehicle classification system. In Proceedings
of the 2001 IEEE Intelligent Transportation Systems, Oakland, CA, USA, 25–29 August 2001. [CrossRef]
7. Harlow, C.; Peng, S. Automatic vehicle classification system with range sensors. Transp. Res. Part C: Emerg. Technol. 2001, 9,
231–247. [CrossRef]
8. Nashashibi, F.; Bargeton, A. Laser-based vehicles tracking and classification using occlusion reasoning and confidence estimation.
In Proceedings of the 2008 IEEE Intelligent Vehicles Symposium, Eindhoven, The Netherlands, 4–6 June 2008; pp. 847–852.
9. Ki, Y.-K.; Baik, D.-K. Vehicle-Classification Algorithm for Single-Loop Detectors Using Neural Networks. IEEE Trans. Veh. Technol.
2006, 55, 1704–1711. [CrossRef]
10. Shokravi, H.; Bakhary, N.; Heidarrezaei, M.; Petr, M. A review on vehicle classification and potential use of smart vehicle-assisted
techniques. Sensors 2020, 20, 3274. [CrossRef]
11. Bhaskar, P.K.; Yong, S.-P. Image processing based vehicle detection and tracking method. In Proceedings of the 2014 International
Conference on Computer and Information Sciences (ICCOINS), Kuala Lumpur, Malaysia, 3–5 June 2014; pp. 1–5.
12. Yu, R.; Wang, G.; Zheng, J.; Wang, H. Urban road traffic condition pattern recognition based on support vector machine. J. Transp.
Syst. Eng. Inf. Technol. 2013, 13, 130–136. [CrossRef]
13. Nanni, L.; Ghidoni, S.; Brahnam, S. Handcrafted vs. non-handcrafted features for computer vision classification. Pattern Recognit.
2017, 71, 158–172. [CrossRef]
14. Fooladgar, F.; Kasaei, S. A survey on indoor RGB-D semantic segmentation: From hand-crafted features to deep convolutional
neural networks. Multimed. Tools Appl. 2019, 79, 4499–4524. [CrossRef]
15. Chang, J.; Wang, L.; Meng, G.; Xiang, S.; Pan, C. Vision-Based Occlusion Handling and Vehicle Classification for Traffic
Surveillance Systems. IEEE Intell. Transp. Syst. Mag. 2018, 10, 80–92. [CrossRef]
16. Zakria; Cai, J.; Deng, J.; Khokhar, M.S.; Aftab, M.U. Vehicle Classification Based on Deep Convolutional Neural Networks Model
for Traffic Surveillance Systems. In Proceedings of the 2018 15th International Computer Conference on Wavelet Active Media
Technology and Information Processing (ICCWAMTIP), Chengdu, China, 14 December 2018; pp. 224–227.
Sensors 2021, 21, 7545 20 of 21

17. Hedeya, M.A.; Eid, A.H.; Abdel-Kader, R.F. A Super-Learner Ensemble of Deep Networks for Vehicle-Type Classification. IEEE
Access 2020, 8, 98266–98280. [CrossRef]
18. Yang, L.; Luo, P.; Loy, C.C.; Tang, X. A large-scale car dataset for fine-grained categorization and verification. In Proceedings of the
2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Boston, MA, USA, 7–12 June 2015; pp. 3973–3981.
19. Krause, J.; Stark, M.; Deng, J.; Fei-Fei, L. 3D Object Representations for Fine-Grained Categorization. In Proceedings of the 2013
IEEE International Conference on Computer Vision Workshops, Sydney, Australia, 2–8 December 2013; pp. 554–561.
20. He, K.M.; Zhang, X.Y.; Ren, S.Q.; Sun, J. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778. [CrossRef]
21. Ng, L.T.; Suandi, S.A.; Teoh, S.S. Vehicle Classification Using Visual Background Extractor and Multi-class Support Vector
Machines. In The 8th International Conference on Robotic, Vision, Signal Processing & Power Applications; Springer: Berlin/Heidelberg,
Germany, 2014.
22. Wen, X.; Shao, L.; Xue, Y.; Fang, W. A rapid learning algorithm for vehicle classification. Inf. Sci. 2015, 295, 395–406. [CrossRef]
23. Matos, F.M.d.S.; de Souza, R.M.C.R. Hierarchical classification of vehicle images using nn with conditional adaptive distance.
In International Conference on Neural Information Processing; Springer: Berlin/Heidelberg, Germany, 2013.
24. Chen, Y.; Qin, G. Video-based vehicle detection and classification in challenging scenarios. Int. J. Smart Sens. Intell. Syst. 2014, 7,
1077–1094. [CrossRef]
25. Cui, Y. Research on Vehicle Recognition in Intelligent Transportation; University of Electronic Science and Technology of China:
Chengdu, China, 2013.
26. Dong, Z.; Wu, Y.; Pei, M.; Jia, Y. Vehicle Type Classification Using a Semisupervised Convolutional Neural Network. IEEE Trans.
Intell. Transp. Syst. 2015, 16, 2247–2256. [CrossRef]
27. Cao, J.; Wang, W.; Wang, X.; Li, C.; Tang, J. End-to-End View-Aware Vehicle Classification via Progressive CNN Learning. In CCF
Chinese Conference on Computer Vision; Springer: Berlin/Heidelberg, Germany, 2017.
28. Jo, S.Y.; Ahn, N.; Lee, Y.; Kang, S.-J. Transfer Learning-based Vehicle Classification. In Proceedings of the 2018 International SoC
Design Conference (ISOCC), Daegu, Korea, 12–15 November 2018; pp. 127–128.
29. Chauhan, M.S.; Singh, A.; Khemka, M.; Prateek, A.; Sen, R. Embedded CNN based vehicle classification and counting in
non-laned road traffic. In Proceedings of the Tenth International Conference on Information and Communication Technologies
and Development, Ahmedabad, India, 4–7 January 2019; ACM Press: New York, NY, USA, 2019; p. 5.
30. Kim, J.; Kim, J.; Jang, G.J.; Lee, M. Fast learning method for convolutional neural networks using extreme learning machine and
its application to lane detection. Neural Netw. 2017, 87, 109–121. [CrossRef]
31. Wang, X.; Zhang, W.; Wu, X.; Xiao, L.; Qian, Y.; Fang, Z. Real-time vehicle type classification with deep convolutional neural
networks. J. Real-Time Image Process. 2017, 16, 5–14. [CrossRef]
32. Tian, Y.; Gelernter, J.; Wang, X.; Chen, W.; Gao, J.; Zhang, Y.; Li, X. Lane marking detection via deep convolutional neural network.
Neurocomputing 2018, 280, 46–55. [CrossRef]
33. Feng, J.; Wu, X.; Zhang, Y. Lane Detection Base on Deep Learning. In Proceedings of the 2018 11th International Symposium on
Computational Intelligence and Design (ISCID), Hangzhou, China, 8–9 December 2018.
34. Van Gansbeke, W.; De Brabandere, B.; Neven, D.; Proesmans, M.; Van Gool, L. End-to-end lane detection through differentiable
least-squares fitting. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, Seoul, Korea,
27 October 2019.
35. Jiang, L.; Li, J.; Zhuo, L.; Zhu, Z. Robust Vehicle Classification Based on the Combination of Deep Features and Handcrafted
Features. In Proceedings of the 2017 IEEE Trustcom/BigDataSE/ICESS, Sydney, Australia, 1–4 August 2017; pp. 859–865.
[CrossRef]
36. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. ImageNet classification with deep convolutional neural networks. Adv. Neural Inf.
Process. Syst. 2012, 25, 1097–1105. [CrossRef]
37. Simonyan, K.; Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv 2014, arXiv:1409.1556.
38. Codevilla, F.; Santana, E.; Lopez, A.; Gaidon, A. Exploring the Limitations of Behavior Cloning for Autonomous Driving. In Pro-
ceedings of the 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Korea, 27 October–3 November 2019.
39. Walambe, R.; Marathe, A.; Kotecha, K. Multiscale Object Detection from Drone Imagery Using Ensemble Transfer Learning.
Drones 2021, 5, 66. [CrossRef]
40. Mikoajczyk, A.; Grochowski, M. Data augmentation for improving deep learning in image classification problem. In Proceedings
of the 2018 International Interdisciplinary PhD Workshop (IIPhDW), Swinoujscie, Poland, 9–12 May 2018.
41. Zhong, Z.; Zheng, L.; Kang, G.; Li, S.; Yang, Y. Random Erasing Data Augmentation. In Proceedings of the AAAI Conference on
Artificial Intelligence, New York, NY, USA, 7–12 February 2020; pp. 13001–13008.
42. Kieffer, B.; Babaie, M.; Kalra, S.; Tizhoosh, H.R. Convolutional neural networks for histopathology image classification: Training
vs. Using pre-trained networks. In Proceedings of the 2017 Seventh International Conference on Image Processing Theory, Tools
and Applications (IPTA), Montreal, QC, Canada, 28 November–1 December 2017.
43. Goyal, M.; Goyal, R.; Lall, B. Learning Activation Functions: A new paradigm of understanding Neural Networks. arXiv 2019,
arXiv:abs/1906.09529.
44. Albawi, S.; Mohammed, T.A.; Al-Zawi, S. Understanding of a convolutional neural network. In Proceedings of the 2017
International Conference on Engineering and Technology (ICET), Antalya, Turkey, 21–23 August 2017.
Sensors 2021, 21, 7545 21 of 21

45. Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.; Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M.; et al.
Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 2015, 115, 211–252. [CrossRef]
46. Chollet, F. Deep Learning with Python; Simon and Schuster: New York, NY, USA, 2017.
47. Deng, J.; Dong, W.; Socher, R.; Li, L.J.; Li, K.; Li, F.F. Imagenet: A Large-Scale Hierarchical Image Database. In Proceedings of the
2009 IEEE Conference on Computer Vision and Pattern Recognition, Miami, FL, USA, 20–25 June 2009.
48. Yosinski, J.; Clune, J.; Bengio, Y.; Lipson, H. How transferable are features in deep neural networks? arXiv 2014, arXiv:1411.1792.
49. Visa, S.; Ramsay, B.; Ralescu, A.; van der Knaap, E. Confusion matrix-based feature selection. MAICS 2011, 710, 120–127.
50. Hay, A. The derivation of global estimates from a confusion matrix. Int. J. Remote Sens. 1988, 9, 1395–1398. [CrossRef]
51. Zhuo, L.; Jiang, L.; Zhu, Z.; Li, J.; Zhang, J.; Long, H. Vehicle classification for large-scale traffic surveillance videos using
Convolutional Neural Networks. Mach. Vis. Appl. 2017, 28, 793–802. [CrossRef]
52. Gao, H.; Cheng, B.; Wang, J.; Li, K.; Zhao, J.; Li, D. Object Classification Using CNN-Based Fusion of Vision and LIDAR in
Autonomous Vehicle Environment. IEEE Trans. Ind. Inform. 2018, 14, 4224–4231. [CrossRef]
53. Shvai, N.; Hasnat, A.; Meicler, A.; Nakib, A. Accurate Classification for Automatic Vehicle-Type Recognition Based on Ensemble
Classifiers. IEEE Trans. Intell. Transp. Syst. 2020, 21, 1288–1297. [CrossRef]