Messages
Proposals
Get Urgent Help In Your Essays, Assignments, Homeworks, Dissertation, Thesis Or Coursework Writing
100% Plagiarism Free Writing - Free Turnitin Report - Professional And Experienced Writers - 24/7 Online Support
Hybrid Automatic Number Plate Recognition Model Using
Cnns And Svms
[Name of the Writer]
[Name of the Institution]
TABLE OF
CONTENTS
2.1.1 Computer Vision techniques used
in Automatic Number Plate Recognition
2.1.2 Machine Learning algorithms used in Automatic
Number Plate Recognition
REFERENCES
The license
plate recognition is a method of recognizing the location of the license plate
on the image that can be made through the method of deep learning or image
processing. The coordinates of the license plate of the vehicle are estimated
at the end of the license plate recognition process
The most
frequently used image processing systems are the automatic license plate
recognition system (ALPR) which are the most effective image processing
systems. These image processing systems are mostly used in public areas such as
the car parks, schools or university car parking as well as in the urban
traffic areas for the detection of number plates to enhance the security in the
city. It used the RFID card as well as the camera images as a solution in the
current ALPR system separately or together
The imager
processing task is complex nowadays because of the current situation that the
ALPR system provides the appropriate outcomes with no errors or with those errors
which are closer to zero. With the widespread use of such technological
systems, the larger amount of data is being obtained nowadays along with new
improvement in the software and the hardware systems. Some limitations and
challenges are existing in these systems through the use of this data, can be
overcome along with the assistance of the deep learning algorithms.
License plate
recognition is significant as well as widespread research problems in computer
vision and image processing. with the latest developments in software and
hardware systems, a large number of related information is being collected.
This has necessitated the development of machine learning ANPR algorithms.
However,
attaining better results of the recognition of License plates is challenging
because of the multiple adverse factors such as lighting changes, occlusion,
and motion blur can lead to large visual variations in plate appearance, which
can severely degrade the performance of the license plate detector. Most of the
existing ANPR machine learning models average accuracy of between 86 %to 90 %
I have therefore proposed a hybrid model that combines CNNs and SMVs to automatically recognize and classify number plates with an accuracy of up to 95%
1) To analyze the different existing Computer
Vision techniques used in ANPR to select the most efficient and appropriate
ones.
2) To analyze the different Machine Learning
methods and choose the most suitable ones for ANPR. To learn how they work and
evaluate their performance based on speed and accuracy metrics.
3) To analyze multiple aspects of CNN to
understand how it works
4) To analyze in detail the working of SVMs and
identified the kernels that are most effective for use with ANPR.
5) To develop an efficient, fast and reliable
hybrid ANPR model using CNNs and SVMs
6) Test and analyze the performance of the proposed model and compare its performance with other existing ANPR machine learning Algorithms
Prototype
development
I intend to
use Anaconda Python 3.7 it is a Python distribution that
comes preinstalled with lots of useful python libraries for data science.
Datasets
BIT-Vehicle
Dataset will be used for this project It includes 10400 labeled vehicle images.
I will divide the dataset as the validation, training as well as the testing
set. The validation set has 1200, training has 8400 as well as the testing set
has 800 images.
Image
processing
Meanwhile, the images in our dataset are
real-world images, some pre-processing methods are essential on the images of
the number plates by using OpenCV. OpenCV (Open Source
Computer Vision) is a library of programming functions that primarily focuses
on real-time computer vision. In simple language, it is the library used for
Image Processing.
Model
Analysis
I will
compare the detection and recognition results from my model with other existing
models
The increase
in the use of the Automatic Number Plate Recognition System has now become the
most researched technology in the past recent years as mentioned in chapter 1.
For this purpose, it has developed various techniques as well as various
methodologies in different computer vision. Therefore, four common stages are
shared by the ANPR methodologies which explained in the next section in this
document along with different related techniques.
Stages of
an ANPR system of
Hybrid Automatic Number Plate Recognition Model Using Cnns And Svms
All Automatic
Number Plate Recognition system is established on some stages which are
mentioned below.
1. Image
acquisition
2. Number
plate extraction
3. Number
plate segmentation
4. Characters
recognition
In the
following lines, these stages are explained one by one along with a review of
the different existing approaches for each of them.
In the stages
of the Automatic number plate recognition is known as the image acquisition as
well as it contains in to capture the videos or the images of different
vehicles running on the road. It can be done in two different ways.
The advantage
is taken by this technique related to the nature of retro-reflective of the
number plate surface. The light is caused by the retro-reflection to reflect
the source unlike the angle or the scatter reflection.
Meanwhile,
the infrared cameras are not available for every person or every vehicle. For
ANPR, the use of HR digital Cameras is extensive. By using this method, the
images are collected which were the common images.
The next
stage of the Automatic Number Plate Recognition is the number plate extraction
in which the number which is given on the number plate is extracted as the
image of the number plate is capture by the cameras. In this stage, many tasks
are performed such as the color extraction, detection of the edges of the
number plate, font or character size on the plate, the position and location of
each character as well as the shape of every character including the background
and character’s colors combination
The most
relatable features of the vehicle number plates are the shapes of the number
plates which are generally come in the rectangular shape. Another feature is to
identify the edges of the rectangle which is a very efficient and common technique
of extracting as well as locating the position of the number of plates. These
methods are very simple and straight forward which are the simplest methods of
the extraction of the number of plates. Therefore, the vehicle number plate’s
edges must be consistent as well as it must not make the images more
complicated. Furthermore, the number of plates must also not have several
unwanted edges that could be disordered with such number plates. This method
also has some significant examples which are: Block based method or the
generalized symmetry transform (GST), Vertical Edge Detection Algorithm (VEDA)
as well as Sobel Filter
The global
image features are being used by another common approach for the extraction of
the number plate. Within the image whose dimension is the same, these
techniques focus on the identification of the connected object to the dimension
of the number plate. The feature is establishing the methods straightforwardly
which are independent where the number plate is positioned in the image which
is captured by the camera. Therefore, the feature can present the issue as well
as they are also considered to consume time with the bad quality images.
Several kinds of examples exist which are: 2D cross-correction or the contour
detection algorithm with the pre-stored number plate template as well as
connected component analysis (CCA)
It is the other type of feature in
the recognition methods that are based on most of those vehicle number plates
sharing similar types of colors. These colors mostly come in the white and
black color combination or the yellow and black color combinations. To
recognize the number of plates, these methods focus on this feature because
these methods analyze the color combinations of number plats in the image. The
method is very effective because it makes the recognition task more efficient
even deformed or inclined the number of plates. Thus, these methods almost
completely dependent on the limitations of the color scheme which is used in the
number of plates. It can be understood with the help of an example. For
instance, the illumination conditions affect the RGB values as well as HLS is
very sensitive to noise. If the number of plate colors is existing, the wrong
detection can take place in other parts of the image. In this section, some
relevant method is also discussed which are related to these methods such as
the HIS color models or the pixel classification based on the HLS, the color
barycentre hexagon model or the segmentation of the color images by mean shift,
as well as the color edge detector.
As everyone
knows, the real center of every number plate is the enrollment number that they
appear, which is a mix of alphanumeric characters. Consequently, techniques
that utilization character highlights exploit this reality to find and
concentrate the number plate from the picture. As a result, these are very
hearty systems that are not influenced by the pivot of the number plate. In any
case, they are generally tedious and wrong discovery can some of the time
happen if extra characters are available in the picture. A few instances of
these strategies are: utilizing calculations for discovering character-like
areas in the picture, doing a flat sweep of the picture or utilizing the width
of the characters.
If all of the
methods which are mentioned in the previous sections already develop better
results unusually at the time of the combined correctly, it can boost the
performance of the system. In simple words, more reliable results are offered
by combining the use of two or more characteristics while it is increased at
the expense within the complex as well as the cost of the computation of the
system. This method has also some important methods which can be the edge
feature or even combining three of them and the combining color as well as the
combining texture features and the color features. Furthermore, the most successful combinations
are those combinations that are explained techniques in previous sections with
the machine learning classifiers
The third
stage of the ANPR system is the segmentation of the number plate which will be
started as the extraction of the number plate as well as it also exists in the
extraction of the number plate in every character of the number plate from the
image of the number plate
pixel
connectivity is the way pixels in an image relate to the pixels that they are
surrounded by, that is, their neighbor pixels. This feature can be employed for
number plate segmentation. This type of techniques is both simple and
straightforward and, indeed, they result to be very robust to number plate
rotation. Nevertheless, they usually fail to extract joined or broken
characters, which can sometimes represent a problem. An example of these
techniques would be: Pixel labeling
Some
different methods and techniques for segmentation of the number plate, are
based on the profiles of projection as well as the extraction of the number
plate image. The running sum of the pixel's values of the image is represented
by these methods in a particular direction which can be vertical or horizontal
mostly. The project profile has the shape of characters which can be utilized
to identify the character’s location or actual position since all of the
vehicle number plates share the collective structure of many alphanumeric
characters in the black color within a row on the yellow or white background,
as it may be detected. These techniques are showing two major benefits. In
other words, those methods are completely self-determining of the location of
the character. Furthermore, the methods can deal with the specific degree of
rotation of the number plate. Although, these methods have also some key
disadvantages which are mentioned in this file. The disadvantages are: the
noise in the image affects the method heavily as well as these methods also
need particular information to extract the alphanumeric or numeric character
within the vehicle number plate because of the variations in different
countries.
Some
techniques that use the specific information of the characters are based on the
idea. If the number of characters as well as the position of those characters
appearing in the number plate are determined already then it is very easy to
extract those characters by using the characters template. In simple words,
those methods establish very simple methods. Furthermore, the particular
information of the format of the number plate limit them as well as the result
of the number plate segmentation can be affected. Some particular examples of
these methods are also mentioned here which are: number plate resizing or the
scanning of number plate within the known template size.
Since the
characters showing up on a number plate, for the most part, have a fixed size,
form displaying of these characters is another method typically utilized for
number plate division. What this sort of method for the most part do is to
estimate the state of the characters to a square shape of fixed measurements
and, at that point, to search for the highlights in the number plate picture
that fit right now, finding the characters. They are exceptionally basic, quick
and clear methods, which can even recognize characters showing up in slanted
number plates. By the by, they can give issues terrible quality pictures, where
a few characters probably won't be very much recognized, prompting an off-base
location of the forms. A few instances of this sort of methods are: Contours
discovering calculations or Shape driven dynamic form model.
Just like in
the previous stage, combining several of the above-described features offers
more reliable segmentation results. But, once again, the price to pay is its
computational complexity. Some examples of this kind of method could be
Adaptive morphology-based segmentation or Dynamic Programming (DP).
Machine
Learning Algorithms for ANPR
There are different types of machine
learning algorithms that are utilized for ANPR. They are classified as follow:
·
Following
the type of problem:
Classification
Problem: There is a
set of samples and each of them corresponds to a specific category or class of
a known discrete set, this type of issue has the objective of assigning the
correct label class to each sample.
Regression
Problem: In general,
this second type seems to tackle the issue of relating the eligible numerical
variables or values to different input samples. Due to it, for example, the
distance to the target based on shape features can be determined.
Clustering
or association problem: The goal of clustering or association issues is all about grouping some
specific instances by the similarity of attributes. Thus, such algorithms can
be applied to the segmentation of an image.
·
Following
the modus operandi:
Supervised:
Generally, this type
of machine learning algorithm is specified or categorized by knowledge
possession of the output during the phase of training, similar to how they
would be learning in the presence of a supervisor. The training data of such
algorithms, in effect, is labeled with a certain value or class so that they
are capable of learning from them and therefore, they can later predict or
determine the corresponding label value or class when they are exposed to a
certain testing set.
Unsupervised:
They are not like the
previous type and these algorithms do not have any type of information of the
put during the phase of training. In simple words, the data associated with
their training is not labeled. In this case, the objective of such algorithms
is concerned with determining unknown or unidentified groupings or patterns of
data by the means of self-learning, which is achieved by them thanks to the
performance or work of an internal self-evaluation.
Every ANPR
system’s final stage is recognizing every number plate character successfully.
Some notable new issues arise at this point such as the noisy or broken
characters, different alphabetical characters for other countries, different
sizes of number plates or the different sizes of fonts including the thickness
of the characters because of some zoom factors. To describe and highlight such
issues, two major techniques based on the use of the two different character
image features exist
All of the
information or data is used by these two methods in the extracted image
character, the operating principle as well as all of the pixel values which
they utilize in the template matching technique. These methods are very simple
and straightforward while these methods can only recognize the non-rotated,
fixed-sized, single font as well as non-broken characters. Therefore, these
methods are made more time consuming than other methods by the fact of
processing all pixels in the image. There are some examples of these methods
exist which are: normalized cross-correction, template matching as well as the
lazy machine learning classifiers.
Using
extracted features
For
recognizing the character, the different types of recognition techniques based
on the idea which is not equally related in the character image, unlike the
previous ones. Although, the use of this method reduces the processing time
significantly as well as the distortion of the image affects a smaller amount
to the outcome. Nevertheless, some additional time is required by the process
of feature extraction as well as the recognition might be degraded by some
non-robust features. There are also some notable examples of these methods
which are mentioned in this document. The examples are Eager machine learning
classifiers based on the feature’s vectors.
2.1.2 Learning algorithms by
Machine used in the Automatic Number plate Recognition:
These days many concepts like
Artificial Intelligence or Machine Learning are seriously thriving almost
everywhere and all this is not happening for any reason because technology like
this in actual is delivering some of the marvelous advances in the area of
computational science. In its effect, Machine learning algorithms at times
offer some of the countless applications out of which Automatic Number Plate
Recognition can also be added. In the subsequent lines, an ephemeral
description for this sort of technology is being given.
Learning is a kind of activity of
getting a knowledge through studying something. In this specific case for the
Machine learning it is a kind of activity of learning that is being carried out
through the computer, but still it enables the construction of different
computerized programs which improves automatically with an experience (Shai Shalev-Shwartz, 2012). There are almost
three different kind of the categories related to the Machine Learning in
accordance to the modus operandi.
Administered: This kind of learning is
being characterized by the knowledge possessing of all the output during a
training phase same like learning from any teacher. In the effect, all the
training data of Supervised Machine Learning algorithm is being labeled out
through the class or any of the certain value to make sure all of them are
capable of learning by them and later on the can predict about the class when
exposed towards the testing set.
Unverified: It is not like the
previous type, they don’t have any knowledge related to the output in complete
training phase. In simple words training data is unlabeled. In this phase, goal
is to know about all of the unknown data patterns or even groupings as well
through the means of self-guided learning.
Fortification of
learning of Hybrid Automatic Number Plate
Recognition Model Using Cnns And Svms:
This is the area of learning. This is
related to taking some suitable actions to exploit the reward in a specific
condition. It has been employed through some machines and even software to find
out the best behavior or the path it needs to take in a particular situation.
This learning is pretty much different from the supervised learning. This learning
is helpful in answering that what kind of data is to be used and will be able
to perform the particular tasks. If the training data set is absent, it has
been bound to learn through the experiences.
At the end, this all is worth
mentioning about some of the most important Machine Learning Algorithms which
are being used for the image processing and that includes support vector
machine SVM, Artificial Neural Networks ANN, K Nearest Neighbors KNN and the
convolutional neural networks. I will here be discussing about the ANPR that
how it is being used along with their limitations.
·
Support
Vector Machine in APNR: This support vector machine is one of those algorithm
which is being used in the APNR system.
1.
While
using SVM in the APNR then developer only gets one output at a certain time.
2.
Classifier
with the SVM have trains on each of the image at specific time.
3.
SVMs
are the kind of models having non-parametric values.
Actually, SVM are a kind of Linear
Classifiers that seem to be from the area of Machine Learning Algorithms which
are supervised plus have an Eager Learning Approach. Considering the points of
group in an area that is from different classes or categories, this type of
algorithms is capable of predicting whether an unidentified or new point
belongs to a class or not. Which the SVM does for achieving this objective is
to find a hyperplane which segregates the points of every class in an optimal
manner. Operation like this is generally the one which helps in assuring the
most of the margin or distance among the nearest points and hyperplane. Due to
it, these algorithms are common referred to as maximum margin classifiers. In
this manner, the vector points which are situated at hyperplane’s area will
belong to the other category, as shown in the figure below. Support vectors
which are carried by these algorithms in the name have usually the subset of
training incidents which are responsible for defining decision boundaries
between classes.
Figure: SVM Operation of Hybrid Automatic Number Plate
Recognition Model Using Cnns And Svms
Generally, for utilizing maximum
margin in provision to the vector machines, it has three different
explanations. Very first reason is that there is a minute error in the location
which gives only a smallest chance of making an issue in classification or even
misclassification. Secondly, the model serves to become resistant to the
elimination of non-support vector points. Thus, the first approximation of what
SMV does is to identify a separative line between data which belongs to two
classes. Actually, SVM is a reliable algorithm which serves to take data as an
input and produces a line which separates those classes if possible as it can
be seen in the figure above. SVM attempts to make or create a decision boundary
in such a manner that the separation between two categories and classes is as
wide as it is possible.
Advantages of SVMof Hybrid Automatic Number Plate Recognition Model Using Cnns And Svms
There are several strengths and
advantages of support vector machines. Some of the major strengths are that
they seem to evade overfitting and they are offer global optimization, which
serves to avoid local minima. A highly relevant aspect is represented by it in
machine learning as well as it makes SVM one of the most reliable classifiers.
There is no doubt that they provide users with exceptional performance on a
wide range of issues, especially two-class issues. However, it is important to
consider that these algorithms have two main negative areas. The first negative
aspect, their extensive requirement of memory and high algorithmic complexity
in large-scale tasks because of the utilization of quadratic programming. In
addition to it, another negative aspect is that insights into kernel parameters
and kernel choice is required
Disadvantages of SVM
of Hybrid Automatic Number Plate Recognition Model Using Cnns And Svms:
Support Vector Machines are determined
for some of the very best classifiers and then they are being used these days
due to some reasons. Still there is not any classifier that is best in their
field and SVM are not the exceptions. Some of the classifiers are much appropriate
and preferable to resolve such issues along with the problems apart from
classifier. Some of the cases are:
• When
quantity of training data and features are very large.
• When
there is a high rate of sparsity and when the majority of features have a zero
value.
Actually, support vector machines
provide different applications in image processing and digital signal.
Convolutional Neural Network (CNN)
This CNN network is from the class of
deep neural networks that is being used most commonly to analyze visual images.
This name also indicates that this whole network works a kind of mathematical
operation that is known as the convolution. This is the special kind of a
linear operation. These are simple neural networks that uses the convolution in
the place of matriculation of general matrix in almost each layer. These are
the networks that were inspired by the biological processes in the different
connectivity methods among the neurons that are responsible for the resembling
an organization for the cortex of an animal. All of these individual cortical
neurons respond towards the stimuli just in any specified area of the visual
field called as receptive field. CNNs use this small processing in contrast to
the classification of image algorithms. All of this means network learn about
the filters that are in the traditional algorithms and were also hand
engineered. This CNN is the network that consists of an input as well as the
output layer, it also consists of many different kind of the hidden layers too.
This hidden layer consists of the different series of the convolutional layer
that has many other dot products. This activation function normally a RELU
layer and then it has many other layers like pooling layers, layers that are
being connected fully along with the normalization layers too that is also
called as referred layer because their inputs as well as the outputs are being
masked away through the activation function along with the final convolution.
All of this has importance for the indices in a matrix that how the weight is
being known at a particular index point.
Architecture of Hybrid Automatic Number Plate
Recognition Model Using Cnns And Svms:
CNN have different architecture as
compared to the basic Neural Network. Basic Neural Networks transforms the
input through putting it from number of different hidden layers of the series.
Each layer is being made from different neuron sets, where individual layer is
being connected to all neurons in a layer that exists previously. Then comes
the last and fully connected layer which is called as an output layer that
helps in predictions.
CNN are slightly different (H. Li Z.
L., 2018). Layers in this network are arranged in three different dimensions
like width, height and depth. Then neurons don’t connect with one another in
the layer or the next one. Final output in this network gets reduced towards
the single vector of the profitability scores that is being arranged along the
complete depth dimension.
CNNs have to different parts:
Hidden layer or the feature extraction
part:
It is that part which performs
different series of the convolutions along with the pooling operations in which
features are being analyzed. Convolution is considered to be one of the major
building blocks of CNN. This term Convolution refers towards the combination of
mathematics in two different ways that helps in the production of third
function. It helps in the merging of two different kind of the information
sets.
In this case of CNN, Convolution is
being performed towards the input of a data by the use of filter or even kernel
to further produce a featured map. At almost every point, multiplication of
matrix is being done and then it sums the complete result on the map of
features.
Same like almost any other part of the
Neural Network, CNN is the one that use the activation function for making the
output completely non-linear (S. Montazzolli, 2019). In the case of CNN, output
will be passed on by the function of an activation. This can be ReLU activation
function. Stride is one of the size of a step of convolution in which the
filter moves every moment. Stride is mostly 1 which means the filter moves
pixel by pixel. Through increasing the size of stride, filter further slides
over the complete input with a large time and it also have less overlap among
the cells.
To keep the size of spatial same,
padding also helps in improvement of a performance and further makes sure that
size of kernel and stride will remain fitted in the point.
After the layer of convolution, it is
much common to add the layer of pooling between the CNN layers. Function of
pooling is to continuously reduce the dimension along with the number of
different parameters. This helps in shortening the training time.
Max pooling is most frequently pooling
that takes up the maximum value in every window. Size of windows need to be
particular. This also decrease the feature size of map and at the same moment
also keep up with the useful data.
Classification part of Hybrid Automatic Number Plate
Recognition Model Using Cnns And Svms:
All of the layers are being connected
fully and then serve the classifier on top of the features that have been
extracted. This helps in assigning the profitability for an object of the image
that what algorithm gives an idea related to it.
After the layers of convolution and
pooling, classification part consists of few layers that are being connected
fully. These layers can only accept the data that is 1 dimensioned. For the
conversion of 3D data into 1D, function flatten in python is being used. This
then arranges the 3D volume into the sector of 1D.
Last layer NN is connected layer
completely. Neurons are arranged in the fully connected form and they also have
full connections towards all of the previous activations in the layers.
Training:
Training out the network of CNN is almost same as the neural network by the use of propagation or the gradient descent. There also lies bit more complexity because of the convolution operation.
CNN or
Convolutional Neural Networks
Generally, whenever there is a
statement about deep learning, CNNs are involved. They are also referred to as
ConvNets and these are the workhorse of the neural network field. They learn to
sort images into some specific categories which is even better than what can be
performed by humans.
Figure:
CNN Recognition of X and O
For understanding how CNNs work, a
simple example will be used: understanding and determining whether an image is
O or X. Actually, this example is sufficient for illustrating the principles
working behind convolutional neural networks.
Figure:
Comparison of Pictures
Generally, CNNs tend to compare
pictures by pieces. In fact, the pieces which are identified by it are referred
to as features. With determining rough feature matches in two images in the
same positions, these neural networks tend to get better at identifying
similarities than whole matching schemes.
Figure: Mini-Image
Each and every feature is similar to a mini-image, which is a tiny 2D array of values. Common aspects of images are matched by features. When X images are considered, the features which consist diagonal lines will match up to the centers and arms of any image of the X.
A. GHAHNAVIEH, e. a. (2018). Enhancing the license
plates character recognition methods by means of SVM. 22nd Iranian
Conference on Electrical Enginering,, (pp. 220-225).
G. R. Gonc¸alves, S. P. (2016). Benchmark for license
plate character segmentation. Journal of Electronic Imaging, vol. 25,
no. 5, pp. 1–5.
Grus, J. (2015). Data Science from Scratch.
Newyork: Jones & Bartlett Learning.
H. Li, e. a. (2018). Reading Car License Plates Using
Deep Neural Networks. International Journal of Computer Applications,
1-8.
H. Li, Z. L. (2018). A convolutional neural network
cascadefor face detection. IEEE, 1-7.
HSU, C. a. (2016). A comparison of methods for
multiclass support vector machines. Neural Networks. IEEE , 1-9.
JOACHIMS, T. (2017). Text Categorization with Support
Vector Machines: Learning with Many Relevant Features. journal of Computer
applications, 1-9.
Jorgensen, H. (2017, june 22). Automatic License
Plate Recognition using Deep Learning Techniques. Norwegian University of
Science and Technology Times, pp. 1-6.
magic, m. (2018). Image Classification :
Step-by-step Classifying Images with Python and Techniques of Computer Vision
and Machine Learning. Wuhan: Wuhan Inc.
Pan, Y., Ge, S. S., Tang, F. R., & Al-Mamun, A.
(2007). Detection of epileptic spike-wave discharges using SVM. . In 2007
IEEE International Conference on Control Applications, 467-472.
Patel, C., Shah, D., & Patel, A. (2013).
Automatic number plate recognition system (anpr): A survey. International
Journal of Computer Applications.
S. Montazzolli, C. J. (2019). Real-time brazilian
license plate detection and rcognation using CNNs. Brazil journal of
science and technlohy, 8.
Sahani, S. S. (2017). License Plate Recognition Using
Convolutional Neural Network. IOSR Journal of Computer Engineering (IOSR-JCE),
19-23.
Sánchez, L. F. (2018). Automatic Number Plate
Recognition (ANPR) System using Machine Learning Techniques. Machine
Learning Techniques.
Schher, R. (2018). Computer Vision Methods for
Fast Image Classification and Retrieval (Studies in Computational
Intelligence). Washington DC: Springe.
Shai Shalev-Shwartz, S. B.-D. (2012). Understanding
Machine Learning: From Theory to Algorithms. Newyork: John Wiley and Sons.
Shi, H. (2019, june 27). An Architecture Combining
Convolutional Neural Network and Support Vector Machine for Image Recognition.
Australian National University press, pp. 3-7.
Toshanlal Meenpal, A. V. (2017). An Overview to
Image Classification . minesotta: Lambert.
V. Kalaichelvi, a. A. (2019). Application of Neural
Networks in Character Recognition. International Journal of Computer
Applications,, 1-9.
Zhao, L. Z. (2018). License Plate Detection with
Shallow and Deep CNNs in complex enviroments. Beijing journal of science
and technology and innovation, 32-37.
Discuss your homework for free! Start chat
