ISSN 2278-3091

Volume 10, No.3, May - June 2021

International Journal of Advanced Trends in Computer Science and Engineering
Available Online at http://www.warse.org/IJATCSE/static/pdf/file/ijatcse611032021.pdf
https://doi.org/10.30534/ijatcse/2021/611032021

An Analysis on Object Recognition Using Convolutional

Neural Networks
Mamoona Saleem1, Salman Afsar2, Ahmed Mateen3, Arslan Zaheer4, Muhammad Tariq5,
Muhammad Asim Raza6
1, 2,3,5,6
Department of Computer Science, University of Agriculture Faisalabad.
4
National University of Computer and Emerging Sciences, Faisalabad
Corresponding Author Email: [email protected]

ABSTRACT
The global development and progress in scientific methods. Due to this reason, a variety of techniques
paraphernalia and technology is the fundamental reason for proposed to elevate the performance of recognition object.
the rapid increase in the data volume. Several significant Convolutional neural network (CNN) among they successful
techniques have been introduced for image processing and model about deep learning, having strong ability of
object detection owing to this advancement. The promising sequential learning features, and the recent research proves
features and transfer learning of Convolutional Neural that the feature extraction done by CNN has a stronger and
Network (CNN) have gained much attention around the reliable ability of discrimination and generalization than
globe by researchers as well as computer vision society, as a hand-crafted features.
result of which, several remarkable breakthroughs were
achieved. This paper comprehensively reviews the data The CNN has attained quite a great success in many areas of
classification, history as well as architecture of CNN and computer vision. An impressive outcome on the ImageNet
well-known techniques by their boons and absurdities. by joining the dataset of Large-Scale Visual Recognition
Finally, a discussion for implementation of CNN over object Challenge (ILSVRC) along with Le Cun's technique and
detection for effectual results based on their critical analysis finally the last fine-tuning techniques to obtain good
and performances is presented. learning. The obtained outcome popularised CNN as it
achieved the error rate of 15 percent vs. 26 percent of
Key words: Neural Networks, Object Recognition, traditional methodology which is overwhelming
Convolutional contribution to the growth of efficient object finding
techniques[6].
1. INTRODUCTION
The volume of image data has been highly increased with In 2014, Zisserman and Simonyan [7] find impact of
the rapid advancements in mobile internet and social media, changing with depth of an CNN on localization as well as
as human beings cannot process efficiently such large organization accuracy in ImageNet challenge, that recovers
volume of data. So, it is expected to handle such tasks then futuristic by utilizingCNN's layers deeper in 16 and 19.
automatically with the aid of automated process. With the A architecture of 16-layer CNN contains five layers of
better understanding of image processing technology, pooling (2x2 neighborhood-max-pooling), fully-connected
extensive recognition of image and exact identification the of three layers and 13 convolutional layer’s (with 3x3 liters).
object target of the image become more and more significant Concealed coatings contain rectified (ReLu) activations. A
[1]. The people are widely concerned about the Fully-connected layers scales channels 4096 to SoftMax
classification of images along with obtaining the semantic 1000 outputs and which can be systemizewith the help of
object category and image location [2], that’s the reason that dropout.
object detection technology has taken wide attraction
globally [3]. Object detection technology tends towards the In 2016 conqueror about the object finding group in the
detection of target objects with the theoretical concepts and ImageNet objective challenge the also based on CNN. This
methods of pattern recognition and image processing, technique has used a amalgamation of CRAFT region
concluding the semantic group objects, and mark targeted scheme generation [8], CNN gated bi-directional[9],
position of target object in image [4]. breakthrough generation, as well as assembling the
clustering.This work has been used so far for the object
It’s a quite challenging taskto recognize image using detection and recognition purposes?
computer technology automatically. Noise disturbance,
complex background, low resolution, attitude and scale 2. RESEARCH METHOD
changes and other factors impacting the object detection The main progress after division is furthermore extraction,
performance seriously. The conventional methods for object representation and the recognition of human actions is
detection were not as robust as convolutional neural unbelievable. When the highlights may contain the data on
networks, as it is not robust to illumination change, thus existence, they are space time volumes (STV) at that point
lacking generalization abilities. Object recognition was quite and when it is possible for us to include discreet Fourier
in-efficient during 2010-2012 in PASCAL VOC changes (DFT) picture contours which spatially contain the
challenge[5], with small gains through the establishment of image power variant. For the extraction of highlights, STV
collection systems and enlists variants using traditional and DFT are all pictured, though they feel distressed. Since
Mamoona Saleem et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1928 – 1935

the nearby highlights are stronger against clamour and 2.3 Data Sources
obstruction. Four electronic databases that are given in Table 1 are
2.1 Identification of the need for review considered as the primary data sources for the extraction of
relevant studies. Moreover, Google scholar had been
In this phase, we look for existing SLRs on Object
considered as the source for external studies. Alt-though,
recognition from Image using convolutional neural
results from Google scholar contains duplicate’s that already
networks. In recent times, many papers are published on
had been extracted from the other four electronic databases.
Object recognition in many disciplines (Image using
So, after removing those du-plicate results only unique
convolutional neural networks Based Techniques). This
results are considered for primary studies.
thing shows that the Image using convolutional neural
networks is becoming popular in upcoming object Search Terms
recognition based on CNN in this table 1. The given search terms had been extracted from the major
terms given which are given in relevant literature studies
Table 1:Electronic Databases and primary question. To find relevant outcomes from the
Identifier Database URL electronic databases listed in Table 1, following search
http://ieeexplore.ieee.or terms were defined. The PICO format of the search terms
ED1 IEEE g/ has been shown in which defines proper categories of the
ED2 ACM http://dl.acm.org/ search terms. Given search terms are combined using the
http://sciencedirect.com conjunction (AND) and disjunction (OR) operators. So, after
ED3 Science Direct / combing the search terms an automated search string has
ED4 Springer Link http://link.springer.com/ been engendered which is given below:
(("Document Title": " machine learning *" OR ",
2.2 Research Questions artificial neural networks" OR" regional proposal") AND
In this review paper, classification of the available image ("Object recognition" OR "convolutional neural networks ")
processing techniques has been elaborated and also detailed AND ("Document Title": "Computer vision “OR "object
literature review of object detection feature of image Detection")
processing has been done (for review of using convolutional
neural networks). There are a couple of important questions In addition, some other keywords had been used to filter out
which will be answered in this review paper. Each inquiry the results from databases; these keywords contain
has appropriate inspiration to demonstrate the need behind broadcasting, aspects, feature, and features. For extraction of
the inquiry in this SLR is these inquiries feature of Object articles from Google Scholar, the search string used is given
recognition from Image using convolutional neural below:
networks. “Object recognition from Image using
convolutional neural networks” This search string is
1) What are the standard image classification reflected as a generic search string and results from the first
techniques? four pages of Google scholar has been extracted which
2) What are other artificial architectures neural contains 20 results for the external category.
network has been used so far for the object
Identification

detection and recognition purposes? Records through DB Additional records

3) Why convolutional neural networks are preferred ACM: 10, IEEE: 52, SD: 4, SL:5 (n = 95)
over other artificial neural network architectures? (n = 69)
4) What is the basic structure of convolutional neural
networks (CNN)?
5) What are the advantages, dis-advantages, variations
Records after duplicates removed
and solutions to previous work proposed in each
(n = 34)
method of convolutional neural networks?
6) Last, what is the best convolutional neural network
Screening

architecture so far considering performance as Records Separated

Records left on the basis of title
benchmark? (n = 100) (n = 20)

To facilitate an organized review about convolutional neural

Records left on the basis of
networks of image processing, its necessary about develop a Records Separated abstract
better understanding of hierarchical classification techniques (n = 85) (n = 15)
which has been described in section 3 below. On the basis of
this taxonomy, literature review of artificial neural
Eligibility

networks, based on object detection, is elaborated in section Full-text articles assessed for Full-text articles left out with
eligibility reasons
4 , the major purpose is to provide a broad overview of (n = 13)
(n = 72)
recent work done including the advances and defects in each
method which related to artificial neural networks. In next
Section 5, detailed review about the convolutional neural
Included

network has done. Comparison at various convolutional Studies included in primary review
neural network methods has done about conclude which is (n = 67)
the best approach so far.
Figure 1: Study Selection Procedure

1929
Mamoona Saleem et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1928 – 1935

2.4 Study Selection Procedure discriminate several textures if provided with triumphing
The study selection procedure of this systematic collected features. Presently, the usage of ANNs in picture processing
works review is visualized in Figure 1. This process of study will increase the aforesaid conventional applications. To
selection consists of 3 phases. Each of these phases is address with low-level image processing duties along with
defined below: image development and noise suppression a part of feed-
forward ANNs and SOMs have they used so far. In order to
Identification deal with image processing of the low level Hopfield ANNs
In this phase, results had been extracted from the electronica had been provided as a tool for finding out an appropriate
databases which are listed in Table 1. These results contain method to deal with complicated (NP-whole) optimization
duplicates, as they are the straight-ahead outcomes that are glitches. therefore, they come to be the correct options to
extracted from the search string. conventional optimization the image processing algorithms
and they could be formulated as optimization issues.
Screening
In this stage, three sorts of filtration happen on the separated Distinct issues addressed within the subject of digital image
outcomes from above stage. Right off the bat, copies of the processing could be described into what we have chosen
outcomes are expelled that are extricated from each about the name of image processing chain (see Fig. 2).
electronic database. After this, results channel based on title. 1) Preprocessing. It has been done initially before
In this progression, the title of the considerable number of applying any other image processing operation. The reason
results are analyzed and just those outcomes are behind is scale-up or scale-down the image as per
incorporated which have a title significant to the point of requirements.
SLR. 2) Data reduction & feature extraction. Extracting them
specific portion or component from an image as per
After these steps, only those results are extracted for the full requirements is called data reduction or feature extraction.
study that has their title as well as abstract relevancy with Extracted features usually have fewer pixels as compare to
the topic of our SLR. the original image.
Eligibility Control 3) Segmentation. Division of image into several regions
that are connected with each other on the basis of some
In this stage, chose considers from the above stage had been
specific criteria. Example is the image operation applied on
full content prepared and just significant articles had been
textures to produce some
chosen for essential investigation and further handling of
4) Object detection and recognition.Observing the exact
our SLR. A legitimate clarification of incorporation and
location i.e. orientation, position, and scales the object
rejection criteria is characterized in the segment underneath
within image.
which unequivocally clarifies the criteria and study
5) Image understanding. It can addresses about the
determination technique based on these criteria.
specific arrangement of object and its in-depth analysis.

2.5Inclusion & Exclusion Criteria Optimization strategies aren't seen as an isolated step in the
The criteria of consideration and prohibition are utilized to whole process. however, it can be considered as a set of
control the qualification for the choice of essential strategies, which help the other steps in following figure 2.
examination from full-content perusing of articles that had
been chosen after the screening stage. It has been seen that Preprocessing
just such articles are chosen for the essential examination
that has legitimate approvals of procedures which are
characterized for extraction of proposal results. Just such
Articles that are recovered from companion checked on Data
diaries and gatherings had been chosen in the last Reduction
investigation. For the choice of an article, it must be
essential that it ought to incorporate curiosity of procedures
and different strategies. Articles from 1988 to 2018 will be
incorporated for study. A few articles that give suggestion Optimisation Segmentation
procedures of mixed media things and tunes and so forth can
be incorporated into the essential investigation. Also, chief
select just such articles which give proposal procedures
dependent on Object recognition from Image using
Obj.
convolutional neural networks. Recognition

2.6Image Processing Algorithms Classification

Conventional Schemes such as the Parzen windows and Image
Bayesian discriminant based on arithmetical pattern Understanding
recognition were famous till early 1990s. Seeing that, ANNs
gained plenty attention and are notably used as a substitute
for clustering techniques and classical pattern recognizing. Figure 2:Architectural process diagram.
Any other attractive trainable machine for object recognition
is non-parametric feed-forward ANNs, which can

1930
Mamoona Saleem et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1928 – 1935

In figure 3, besides, task achieved using algorithm, its can Various architectures of network have been introduced to
be processing competencies are in part regulate via the deal with the object variations position and orientation. An
distraction degree about the enter records. Following approach has been taken that is proportional to illumination
abstraction stages has been distinguished based on change. In order to gain appropriate results and for better
numbering. classification, the distinctionis required between invariant
recognition in 2D&3D images. A novel approach has been
introduced for object recognition knows as what-and-where
2. Local feature filter that was proportional to the 2D translations, scale, and
1. Pixel level 3. Edge level
level in-plane rotational. Its can be combination of what (filter
bank) along the other with where (an invariant module).
Several other methodologies need to learn via explicitly
training. Objects statistical intensity model was built by
Egmont-Peterson and Arts. A two-stage ANN approach so
6. Scene 5. Object set
far for recognition of nodules chest radiographs was
4. Object level developed by Penedo et al. Nodule sub-images were partly
characterization level
used to trained these ANNs.

Figure 3: Abstraction Level of Image Processing Chain Rare cases like object detection occlusion or the existence of
more than one objects inside the image are processed by a
classifier that is rarely been taken into consideration.
Image Processing Based on Neural Networks McQuiod developed an experimental structure which is
In this section, we will be review various object detection capable for multiple object detection simultaneously within
systematic techniques strut on artificial neural networks an image[9].
before moving to convolutional neural networks (CNN)[10] Convolutional Neural Network
that contains the review of feature based other artificial
neural networks too for object detection. Before starting discussion on various aspects of
convolutional neural networks, it is necessary to mention
Object recognition here other artificial neural-network approaches based on
Object recognition is among the important issues based on feature-based object detection including: feed-forward
computer vision and is particularly a complicated problem ANNs, Hopfield ANNs, a fuzzy-ANN and RAM-based
to accomplish. In lots of respects, object recognition is ANNs. SOMs are often executedin order to perform feature
pretty a similar other to computer vision functionality, it as extraction before object recognition, even if SOMs are
consists of creating an package that’s forms a regular trained to perform object classification.
behavior to deformation and adjustments in the perspective
in addition to lighting fixtures. A prime factor that makes Compared to the pixel-based approach described in the
object recognition a distinct issue is it includes each and previous section, neural architectures are developed on a
classifying and finding regions of an image [11]. smaller scale for feature-based object recognition. It reflects
the point that foremost focus to develop or select them most
Through late 2000s, predominant answers for object optimal structures for them recognition tasks. Various
recognition is to make use of feature descriptors, which feature-based approaches have a common pathway that
includes scale-invariant function transform (SIFT) discrepancies in rotation and scale are crumbled to the
[12]which evolved through David Lowe(1999) along with structures, e.g., statistical moments. It is also noted that a
histograms orientated gradients (HOG) [13]which became certain measure of noise always influences the computed
popularized by 2005. By 2010s, technology has been moved structures, as a result of which the recognition performance
towards the usage of convolutional neural networks[14]. is deteriorated. Therefore, the subsequent classifier performs
the major task of filtering out the noise and distortions due
Pixel-based object detection to these features. Additionally, whenever a large object is to
Several ANN methods have been introduced for object be detected and densely sampled, feature extraction should
recognition based on pixel data. Strategies using the weight be performed. Contrarily, a neural classifier is comprised of
sharing algorithms[15], the recurrent networks developed in so many parameters that a qualified generalization will be
1980, the ART network is the contribution of Grossberg, obstructed.
mixture-of-experts used to divide the space problem into
homogeneous regions, fuzzy ANNs consists of fuzzy logics The major issue in solving the computer-vision, image
and ANNs method, bi-directional auto-associative memories processing-based issues using a conventional neural network
(BAM) is a sub-class of recurrent neural network which was is that a usual normal image comprises of a large quantity of
brought by the Neocognitron is a hierarchical, multilayered information. A monochrome low-resolution image i.e.
ANN delivered with the aid of piecewise-linear neural (620x480) contains about 297-600 pixels. A general
classifiers based at the Kohonen learning vector quantization assumption leads to a conclusion that if every pixel of this
(LQV2) , higher-order ANNs and Hopfield ANNs. A picture is allocated a separate inter-linked, 297-600 weights
hardware has been designed to cope with object detection required for each neuron. Must need 2,073,600 weights for a
problem based on ANNs: the RAM network suited full HD image (1920x1080) andIf the pictures provided are
particularly to implement WSI and the optical polychrome, then quantity weights about increases
implementations. ultimately, Self-Organizing Feature Maps according to them color channels (typically three). Thus, it
(SoMs) being used to extract features from pixel-based data. can be seen that the general sum of free factors can be in the

1931
Mamoona Saleem et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1928 – 1935

link rapidly turn into particularly enormous as the Pooling Layers

doppelgänger dimensions’ rise. Moreover, big data effects The main responsibility of pooling layers is to decrease the
overfitting and in-efficient enactment. spatial resolution of feature maps. Basically, network deep
layers require less information about features exact spatial
Moreover, many patterns finding methods involves that locations, at the same time more filter matrices are
explanation is translationally constant. It’s can wasteful for requiredin order to recognize multiple high-level patterns.
the separately train of the neurons observing that the similar The depth of data volume can be increased by decreasing
systematic technique can be the right-bottom corner and the height and width of data-volume that could be helpful in
left-top crook of an image. A fully-connected neural keeping the computation time at a reasonable level.
network failed to notice this structure and therefore failed in
this case, thus in such scenarios. Data volume size can be reduced by adding a pooling layer
after a convolutional layer. The layer down-samples the
Generalized Architectural Overview activation maps. The main issue with pooling layer that it
As CNN's are feedforward networks, so the information may destroy the information about spatial relationships b/w
flow only occurs in one direction, from input to output. subparts of patterns. They can do by adding a max pooling
CNN's are biologically inspired just like an Artificial Neural layer after infinite and convolutional layer [17].
Networks (ANN). Visual Cortex in brain comprises of
alternating layers of simple and complex cells (Hubel & Fully Connected Layers
Wiesel), motivates their architecture. CNN came in various Multiple pooling and convolutional layers are bonded
variations but in general, it consists of convolutional and together at the top of each order in the other to extract more
pooling (or subsampling) layers, which are grouped into abstract article moving concluded the network. Every
components. These connected layers follow these neuron in one layer connects to another neuron in each
components. Components are often bonded together at the layers using full connected layer phenomenon . It is in
top of each other, thus forming a deep model in figure 4. standard of the same as the outmoded multi-layer perceptron
neural network.
Input image Training
Artificial Neural Networks uses learning algorithms to
adjust their free parameters (i.e., the biases and weights) in
order to get desired network output. Commonly used
Convolutional
layers algorithm for this purpose is backpropagation [18]. It
proposes the solution to iterate weights to reach at constant
point. In old version, gradient descent method has been used
Fully
for optimization purpose. However, in modern literature,
connected gradient descent is undoubtedly a time-consuming technique
layer and found unreliable to for minimization of errors.
R-CNN
In this section, we will study various methods that
Output class joinCNN's with regional proposal classification along with
how are they generated [19] (also called as Region of
interest RoI)
Figure 4:CNN Image Classification Pipeline [16]
Overview/Description
R-CNN forward computing comprises of various stages, as
Convolutional Layer shown in fig. 5. After taking images as the input, at first, the
This layer acts as a feature extractor, themain responsibility region of interest (RoI) has been generated [20]. Generated
is to learn the feature representations of their input images. RoIs are category-independent bounding boxes and they
Convolutional layers comprise of neurons which are have a high ratio of comprising the targeted object. A
arranged in feature maps having a sympathetic field, which distinct method called Selective Search has been used for
can be connected to the neighborhood about the neurons of making the RoIs (see a reference for appropriate data).
the backward layers through the set of trainable weights.
Inputs are convolution using learned about the weights of It has been used to determine features from every region
the compute a featured map, obtained convolved results sent proposal also called Region of Interest (RoI). The sub-image
via non-linear activation. Neurons inside a feature map have restricted inside the bounding-box has been used toward
weights that constrained to be equal; but, diverse article parallel the input size of the image in CNN and then passed
maps inside the similar layers may have different weights In further near the network. Once network can be successfully
order to compute the kth output feature map,Ykcan be extracted structures from the input, the structures are then
computed as: passed to support vector machines (SVM) thats gives them
Yk = f (Wk∗ x) (1) final classification of vectors.
Where x is the input image; Convolutional filter belongs to
kth feature map is represented by Wk. the * sign used to
calculate the product of the filter model at each point of the
input image.

1932
Mamoona Saleem et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1928 – 1935

input along with identification of Region of Interest (RoI).

Fully Image has been processing using various layers comprises
Convolutional
Input image connected Output class
layers of convolutional layer as well as max pooling layer in order
layer
to generate a convolutional article map. Next step is to
extract an article vector from each Region of Interest (RoI)
Figure 5:Stages of R-CNN Forward Computation. using article map. Then, feature vector will be provided as
an input to the fully-connected layers, that further acts as an
In the above figure 5 this design has been prepared in input to further two output layers, that aresoftmax layer,
multiple phases, begins off advanced with the convolution responsible for producing probability estimation and the
network. When CNN learning has been finished, help to later known as bounding box (responsible for refinements of
vector machines (SVM) are then passed further to CNN initial candidate boxes)[21].
features. Eventually, the learning procedure begins for
region proposal. Classification evaluation
As described via the authors, fast R-CNN takes quite a short
Issues
span of time in-order to classify an image in comparison to
R-CNN is quite a significant architecture because it conventional R-CNN, nearly takes not more than a second
contributed toward provision of a first working solution for time using state-of-art Graphical Processing Unit. The
object recognition using CNN's. As R-CNN is among the motive at the back of is because each RoI has used same
first one, it has a number of defects that have been addressed feature map.
by later research.
As recognition time decreases, overall computation time
In 2015, Girshick pointed out 3 main problems of R-CNN in starts depending on the response time of region proposal
his paper: generation method. Due to this, Region of Interest
- The training comprises of couple of steps, as shown in production forms a computational barrier. In case of
fig. 5 multiple RoIs, the evaluation of time span consumes on the
- Secondly, the training takes quite a long time which is convolutional layer dominates the period expended
making it quite expensive for practical operational estimating the fully-connected layer. Time consumption
work. Also, for both Support Vector Machine as well as canister be decreased by compressing the connected layers
region proposal training, features have been taken and implementing truncated singular value decomposition but
stored on disk from each region proposal. A long period the drawback is it will result into a bit of issue in accuracy
of computation is required along quite a large number but overall performance gain with respect to time will be
of disk space. more then 30 % of original time.
- Last one and the quite important, processing of object
detection algorithm is quite slow and time consuming Training
for the image even using the GPU. Forward computing As described [16], Fast R-CNN is quite more efficient to use
is one of the reason along with only one Regional for training purposes rather than R-CNN, through almost
Proposal can be generated at a time. In case of multiple nine-fold decrease popular training time-span. The can be
RoIs, system is not capable of handling them whole network (including the fully connected layers and RoI
concurrently thus due to which an overlapping is layer) could be trained by using the two algorithms that are
occurred in results, which make the method non-reliable back-propagation and stochastic ascent descent. Usually,
and in-efficient. pre-trained algorithm can be chosen as a starting opinion
and at that point refined to next-level. Mini-batches of N-
Fast R-CNN
images is used for training. Each mini-batch image is
This architecture was, moving toward more practicality, this responsible for providing a sample to R/N RoIs. If the
architectural method for object recognition focused towards connexion over amalgamation with the. ground-truth
performing the forward permit about the CNN consists of container is above 0.5, then RoI samples are consigned to
complete image, as a substitute for the isolation Regional the class. Beside this, left over RoI belongs to the
Proposal. background class.

As per classification, computational and memory usage have

been shared for the RoI from the same image. The original
image has been tossed horizontally with a probability about
0.5 for data-augmentation. Using a multi-task loss function,
the bounding box regressors and the soft-max classifiers are
fine-tuned together, both are considered to be the accurate
class of the experimented RoI and offset about sample
bounding container from accurate bounding box.

3. MAIN RESULTS
They can be main Three results support the Fast R-CNN
Figure 6: (Fast) Region-based Convolutional Networks (R-CNN). contributions:
General description - State-of-the-art mAP on VOC07, 2010, and 2012
- Fast training and testing matched to R-CNN, SPPnet
Figure 6 explains the basic architecture of fast R-CNN. This
(Fast R-CNN) technique receipts entire image by way of an
1933
Mamoona Saleem et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1928 – 1935

- Fine-tuning convolution layers in VGG16 improves effective solution to numerous computer vision problems.
mAP Afterward, the evolution about the convolutional object
detection from the R-CNN can be recently introduced
optimal techniques is demonstrated. This review not only
3.1 Faster R-CNN
describes the advancement in the structure of the CNN but
A Faster R-CNN network is presented and trained for RoI its implementation and computational effectiveness are also
production and its detection. Primarily, training has been illustrated. In fact, the paper describes the comprehensive
done separately on networks. Afterward, the next step review of techniques to overcome computational
consists of the combining and refining of networks. While bottlenecks, integration of different phases into the CNN,
refining process, certain layers are kept fixed due to which improvement in response time, automatic error correction
many layers get trained in return. and its optimization over time lapse.

A single image is provided as an input to the trained The Future Work

network. Feature maps generated from the image using The exact study of the designed techniques, implementation
shared fully convolutional layers. Region Proposal and real-time hardware tests can be a future trend for the
Network(RPN) taking feature maps as an input. RPN is researchers. For further research response, time reduction on
considered as a "weak" detector, only responsible for commercial computers would be a thought-provoking topic.
detection Whether there is an object and generate a Due to hardware cost, the claim of real-time performance by
bounding box proposal Object. But in fact, the convolutional various techniques is not achieved yet. It can be achieved by
layer is "powerful" enough to detect, locate, and classify implementing these techniques on real-time hardware. So
objects simultaneously [22]. RPN then produces the region that more and more applications can be commercialized for
proposals, which are fed as an input along with article maps consumers. Hardware cost and size reduction can also be an
into the final recognition layers. Detection of the layers optimal approach towards future trends. Moreover,
containRoI pooling layer ot the as a result of which final according to aforementioned suggestion, an approach can be
classifications. acknowledged for producing a complete convolutional or
neural system, which can be able to learn inherent features
A foremost reason for using shared convolutional layers is for an object classification automatically. Correspondingly,
the negligible computational cost of region proposals. this system should also be capable of differentiating the
Moreover, computing region proposals using CNN has the object can be since a convinced part of a scene. However,
extra benefit of its dependency on GPU as compared to old- this systemic technique can be implemented by integrating
fashioned RoI generation methods (i.e. Selective Search), geometric interference and CNN. Yet, it depends on time
which were applied using CPU. and scope that which research direction the research takes.

REFERENCES
3.2 SSD
The Single Shot MultiBox Detector (SSD)[23] is a unified [1] C. Szegedy, A. Toshev, and D. Erhan, Advances in
framework for the object detection with a single network. It Neural Information Processing Systems,Deep
often promotes integrated detection. Neither, it produces any Neural Networks for object detection, vol. 26, pp.
proposal nor it takes part in resampling the image segments. 2553–2561, 2013.
However, the process of object detection in SSD is carried [2] K. Q. Huang, W. Q. Ren, and T. N. Tan, A review
out through a single pass of CNN[24]. on image object classification and detection,
Chinese Journal of Computers, 2014.
It somehow resembles a descending window method, in [3] F. Xie, M. Zhang, J. Zhao, J. Yang, Y. Liu, and X.
which procedure is instantiated with an avoidance of the Yuan, A Robust License Plate Detection and
regular bounding values. Diverse scales and aspect ratios Character Recognition Algorithm Based on a
can be included in it. The calculations for the object Combined Feature Extraction Model and BPNN,
predictions depends on the boxes, which includes offset Journal of Advanced Transportation, vol. 2018,
parameters and hence, responsible for predicting the 2018, doi: 10.1155/2018/6737314.
correctness of bounding box covering the object as [4] O. Russakovsky, J. Denf, S. Hao, K. Jonathan, S.
compared to a defaulted box[25]. Satheesh,M. Sean and H. Zhiheng, ImageNet Large
This algorithm is based on several scales which uses the Scale Visual Recognition Challenge, International
feature maps from different convolutional layers as a Journal of Computer Vision, vol. 115, no. 3, pp.
response to the classifier[26]. As this method generates quite 211–252, 2015.
a large number about the classifier, boxes filters the boxes [5] D. Hoiem, D. S.K., and J. H. Hays, Pascal VOC
up to a suppression stage, as a result of which the boxes 2008 Challenge, World Literature Today, 2009.
below a certain threshold are eliminated. [6] T. Y. Lin et al., Microsoft COCO: Common
Objects in Context, Springer International
4. CONCLUSION Publishing, vol. 8693, pp. 740–755, 2014.
[7] K. Simonyan and A. Zisserman, Very deep
This paper has presented a detailed review of the convolutional networks for large-scale image
fundamental context for CNN implementation and recognition, arXiv preprint arXiv, p. 1409.1556,
classification for the object detection phenomenon based on 2014.
its pros. and cons. Moreover, the constraints of conventional [8] B. Yang, J. Yan, Z. Lei, and S. Z. Li, Craft objects
neural networks in image recognition are demonstrated. The from images, In Proceedings of the IEEE
paper also describes the advantages of the CNN as an Conference on Computer Vision and Pattern

1934
Mamoona Saleem et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(3), May - June 2021, 1928 – 1935

Recognition, pp. 6043–6051, 2016. 2017.

[9] X. Zeng, Zeng, Xingyu, W. Ouyang, Y. Junjie, Li, [23] W. Liu, D. Anguelov, E. Dumitru, C. Szegedy, R.
Hongsheng, X. Tong, K. Wang, Y. Liu,Crafting Scott, F. Cheng-Yang and C. B. AlexanderSsd:
gbd-net for object detection, arXiv preprint arXiv, Single shot multibox detector, In European
pp. 1610–2579, 2016. Conference on Computer Vision Springer, pp. 21–
[10] M. Egmont-Petersen, U. Schreiner, and S. C. 37, 2016.
Tromp, Detection of leukocytes in contact with [24] M. Srivastava, S. Mishra, H. Singh, and H. Singh,
the vessel wall from in vivo microscope HIERARCHAL CLASSIFICATION OF
recordings using a neural network, IEEE Trans. SATELLITE, vol. 7, no. 19, pp. 2524–2529, 2020.
Biomed. Eng, vol. 47, no. 7, pp. 941–951, 2000. [25] T. A. Al-asadi and M. A. Almaamory, HIDING
[11] R. Girshick, Fast r-cnn, In Proceedings of the IEEE FINGERPRINT IN IRIS IMAGE BASED ON,
International Conference on Computer Vision, pp. vol. 7, no. 19, pp. 3902–3909, 2020.
1440–1448, 2015. [26] B. Mallikeswari and D. P. Sripriya, A DOUBLE
[12] D. G. Lowe, Object recognition from local scale- FILTERING-DENSITY ALGORITHM FOR
invariant features, In Computer vision, 1999. The ENHANCING IMAGES OF DIFFERENT
proceedings of the seventh IEEE international TYPES, JCR, vol. 7, no. 13, pp. 2394–5125, 2020,
conference, vol. 2, pp. 1150–1157, 1999. doi: https://doi.org/10.1109/cvpr.2016.232.
[13] N. Dalal and B. Triggs, Histograms of oriented
gradients for human detection. In Computer
Vision and Pattern Recognition, CVPR, IEEE
Computer Society Conference on (2005), vol. 1, pp.
886–893, 2005.
[14] R. Girshick, J. Donahue, T. Darrell, and J. Malik,
Rich feature hierarchies for accurate object
detection and semantic segmentation, In
Proceedings of the IEEE conference on computer
vision and pattern recognition, pp. 580–587, 2014.
[15] S. Ren, K. He, R. Girshick, and J. Sun, Towards
real-time object detection with region proposal
networks, In Advances in neural information
processing systems, pp. 91–99.
[16] X. Xinfeng, D. Du, L. Qian, Y. Liang, W. Tang, L.
O. Zhong, L. Mian, P. H. Huynh and R. S.
Monggoh. Exploiting Sparsity to Accelerate Fully
Connected Layers of CNN-Based Applications on
Mobile SoCs, ACM Transactions on Embedded
Computing Systems (TECS), vol. 17, no. 2, p. 37,
2018.
[17] J. Wang and X. Xiaolong, Non-local neural
networks, The IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), vol. 1, no.
3, 2018.
[18] C. M. Bishop, Pattern Recognition and Machine
Learning (Information Science and Statistics),
Springer-Verlag New York, Inc., Secaucus, NJ, USA,
2006.
[19] A. Krizhevsky, I. . Sutskever, and G. E. Hinton,
Imagenet classification with deep convolutional
neural networks, In Advances in neural
information processing systems, pp. 1097–1105,
2012.
[20] K. E. Van de Sande, J. R. Uijlings, T. Gevers, and
A. W. Smeulders, Segmentation as selective
search for object recognition, In Computer Vision
(ICCV), IEEE International Conference, pp. 1879–
1886, 2011.
[21] Z. He, D. Liang, S. Zhang, X. . Huang, and S. Hu,
Traffic-sign detection and classification in the
wild, onference on Computer Vision and Pattern
Recognition (CVPR).
[22] Ren and Shaoqing, Faster R-CNN: towards real-
time object detection with region proposal
networks, IEEE Transactions on Pattern Analysis
& Machine Intelligence, vol. 6, pp. 1137–1149,

1935