International Journal of Computer Applications (0975 – 8887)

Volume 177 – No.3, November 2017

Design of a Recognition System Automatic Vehicle

License Plate through a Convolution Neural Network

P. Rajendra K. Sudheer Kumar Rahul Boadh

Department of Mathematics, Department of Computer School of Basic and Applied
Madanapalle Institute of Science & Engineering, KG Sciences,
Technology and Science, Reddy College of Engineering & K.R. Mangalam University,
Madanapalle, India Technology, Hyderabad, India Gurgaon, India

ABSTRACT
The present work is a study on the practical application of registrations from images captured by a camera, and are
Learning process (Deep Learning) in the development of a mainly used in monitoring and control devices. In these
system of Automatic recognition of vehicle license plates. systems the recognition of the images of numbers and letters
These systems commonly referred to as ALPR (Automatic can be implemented through different Machine learning,
License Plate Recognition) - are able to recognize the with the most common being the neural networks of type
content of vehicles from the images captured by a camera. Multilayer Perceptron (MLP) and Support Vector Machines
The system proposed in this work is based on an image (SVM). Even some systems choose to use commercial
classifier developed through supervised learning techniques Optical Character Recognition (OCR) for this purpose.
with convolution neural network. These networks are one of
the most profound learning architectures and are specifically 1.3 Classification of images
designed to solve artificial vision, such as pattern The recognition or classification of images consists of
recognition and classification of images. This paper also assigning an image a label of a defined set of categories
examines basic processing techniques and Image based on their characteristics. Despite seeming to be a
segmentation - such as smoothing filters, contour detection - relatively trivial problem from our perspective, is one of the
necessary for the proposed system to be able to extract the most important challenges artificial vision systems. Factors
contents of the license plates for further analysis and such as scale, lighting, deformations or partial concealment
classification. This paper demonstrates the feasibility of an of objects make the classification of images is a complex
ALPR system based on a convolution neural network, noting task, to which a great deal of effort to develop sophisticated
the critical importance it has to design a network architecture pattern recognition techniques, which do not always produce
and training data set appropriate to the problem to be solved. the expected results. From the point of view of machine
learning, the classification of images is a supervised
General Terms learning problem, in which the classifiers algorithms
Deep Learning, Tensor flow, Python generate a model from a dataset or set of previously
categorized images. The model obtained is used later to
classify new images.
Keywords
Convolution Neural Network, Deep Learning, ALPR

1. INTRODUCTION
1.1 Artificial vision systems
Artificial vision is a branch of artificial intelligence whose
purpose is design computer systems capable of
"understanding" the elements and characteristics of a scene
or image of the real world. These systems allow extracting
information - numerical and symbolic - from of the
recognition of objects and structures present in the image.
Artificial vision is closely related to Image processing and Fig.1: Digital image expressed as an array (pixel values are
pattern recognition. In earlier, used to facilitate the Fictitious)
localization and detection of areas of interest in the images; For these algorithms the images are three - dimensional
latter are used to identify and classify objects and structures matrices whose dimensions are the width, height and color
detected according to their characteristics. depth. The content is being of each position of the matrix a
numerical value representing the intensity of each pixel in
1.2 ALPR systems the digital image.
Systems ANPR (Automatic License Plate Recognition)
vehicles are a particular case of systems of artificial vision. Until the popularization of convolution neural networks
These systems are designed to "read" the content vehicle (CNN), the systems classifiers based on support vector

47
 International Journal of Computer Applications (0975 – 8887)
Volume 177 – No.3, November 2017

machines (SVM) were the ones that presented the best necessary develop selection processes and attribute
results in recognition problems and classification of images. extraction. The learning algorithm of these networks allows
The main drawback of these systems is that they require a extracting the attributes or characteristics of each class from
prior process of extracting the relevant features of the a set of training data previously classified. These attributes
Images, almost always designed to the problem of the are the weights of the different neurons in the network, and
intended purpose solve, for which sophisticated techniques their values are calculated iteratively by a supervised
of detection of objects are used and patterns - histograms of learning method called back propagation or "Backward
oriented gradients (HOG),SIFT descriptors etc. The SVM error propagation".
classifier algorithm is trained from the characteristics drawn
to a subset of the training dataset, so the efficiency of the In broad outline, the algorithm consists of two repeating
generated model depends on how representative they are. stages iteratively for each element of the training set:
The most negative aspect of these systems is that they really • In the first class to which it belongs is calculated input
do not learn the characteristics or attributes of each copy according to the current values of the network weights.
category, as they predefined in the extraction step. In Once classified, the algorithm determines the validity of
addition, these systems are extremely sensitive to variations such classification by an error function that calculates cost
in scale, lighting, perspective etc. or how good or bad it is, comparing it with the class to
which the example of training introduced into the network.
1.4 Artificial neural networks • Known error, the second stage of the algorithm propagates
Artificial neural networks are an automatic learning back to all neurons in the network that have contributed to
paradigm inspired by the functioning of the biological brain. the classification of example, each receiving the "portion" of
These networks are composed of interconnected neurons that the corresponding error in function of their contribution, so
collaborate to produce an output from the input data of the that they update the weights proportionally, so that the new
network. Each artificial neuron or perceptron is a processing values reduce the classification error.
unit receives a series of input signals multiplying by a given
weight (Synaptic weights). The neuron calculates the sum of Gradient descent method allows calculating a minimum
the product of each input by its corresponding weight - to value of the error function is obtained. To calculate this
which a correction factor is usually added or bias - and value, the algorithm descends iteratively by the slope
applies the resulting value to an activation function that (Gradient) from the current value, forward in each iteration
produces a output value or other, depending on whether the step with a length or rate specific learning. The main
sum of signals and weights exceeds one certain threshold. drawback of the neural networks described in this section is
the difficulties posed by the total interconnection between
layers of neurons with increasing dimensionality of the data
of entry. For example: a color image of 300x300 pixels
requires a input layer with 270,000 pesos (300x300x3), if
multiple layers are added intermediate network to the
number of weights and bias necessary to grow exorbitantly,
increasing the computational cost and the risk of over
network training (over fitting).

1.5 Convolution neural networks

The convolution neural network (CNN) is a particular type
of neural network. Neuronal system inspired by the
functioning of the visual cortex of the brain. These networks
are designed to solve artificial vision problems such as
recognition of patterns, although they may have other uses
Fig.2: Schematic an artificial neuron with ‘m’ inputs such as classification of texts or processing of natural
language. Like the "conventional" neural networks, they
Artificial neural networks are organized in layers of neurons receive an input that transform through a series of layers of
where each layer processes the information received from neurons, but in this case the input is an image represented as
the previous layer. The number of layers and the type of a three-dimensional matrix - Width, height and color depth -
neuron activation function determines the complexity of the containing the numerical values of the Pixels. The layers of
problems that the network can solve: from detecting simple neurons are also organized in a way three-dimensional, and
patterns in linearly separable data (one layer); to complex the result of the different transformations carried out on the
nonlinear relationships between the input data (more than net is the class or category to which that image belongs. The
three layers). Networks with one or more intermediate layers convolution networks are constructed using four types of
between the input and the output are which they are called layers: input layer, convolution layer, layer pooling and
deep networks, and are the basis of Deep Learning. In the fully connected (FC, Fully Connected). Convolution layers
following diagram you can see a network with a Hidden in neurons are not fully interconnected, but rather with a
layer, it is also observed how in these networks the neurons small region of the anterior layer. In addition, this layer
of a layer are fully interconnected with those of the next. The neurons share the same weights and bias, which greatly
great advantage of neural networks over other methods is reduces the number of network parameters. Graphically,
their machine learning mechanism, through which it is not these clusters can be interpreted as a window that runs the

48
 International Journal of Computer Applications (0975 – 8887)
Volume 177 – No.3, November 2017

image or input layer, sliding from left to right and up and system will execute the segmentation it is consisting
down until you reach the end. The size of the sale and its detecting possible numbers and letters present on the license
length displacement are parameters to be determined plate. The code module that performs this process leads to a
according to the characteristics of the problem. As the segmentation based on the detection of contours, using it
window moves, the corresponding neurons in the features the popular OpenCV computer vision library. The
convolution layer calculate the product between the matrix stage character recognition will be made through a CNN
of shared weights, bias and the pixel values of the region to previously trained with a set of data with computer
which are connected, sending the result of this operation to generated numbers and letters with different typographies
an activation function determined - usually RELU(Rectified and styles. The output of the system will be a string
Linear Unit) function, which returns the max (0, x)value. composed of the characters detected in the image of the
The result of these calculations is a series of maps entry plate.
activation allow detecting the presence of a certain
characteristic in the image of entry. This feature is defined 3. DESCRIPTION OF THE DATA
by the weight matrix shared and the value of bias; the
combination of both is called filter or kernel and its 3.1 Training data set
application to a region of pixels is called convolution. As The effectiveness of a network-based image classification
you can see in the previous image, in the convolution layers system CNN depends on both the architecture and network
it is several filters - weight matrices - to detect more of a settings, and of the data set used for their training and must
feature in the picture. The convolution layers are usually sufficiently broad and representative of the problem that is
accompanied by a layer of pooling which condenses the wanted solve. In order to train the neural network of the
information collected, leaving only the maximum values of proposed system, the set Chars74k data created by T. M.
an area of the convolution layer, which dimensions of the Campos and Varma [3]. This dataset has over 74,000 images
input volume for the next layer. The typical architecture of numbers and letters in PNG format organized in three
a CNN network is a succession of layers pooling and collections: characters manuscripts, drawn characters and
convolution, being habitual that the last layer is a Fully photographs of everyday scenes computer generated
Connected network which is responsible for calculating the characters. This latest collection has more than 60,000
scores obtained by the input image for each of the classes or images in grayscale of digits and letters represented with
categories defined in the problem. The learning algorithm different fonts and styles (normal, bold and italics), a priori
of these networks is seen in back propagation gradient suitable for the CNN network "learn" to recognize patterns
descent and other similar methods. But in these networks of different numbers and letters of the license plate.
training aims at the different layers of neurons learn the
filters or characteristics of low and high level (lines, edges,
etc.) which represent each class or category of problem
image. CNN networks are today the "state of the art" in the
field of vision Artificial, where they have demonstrated a
superior efficiency to other techniques, and this is due to
three key advantages of the architecture:

 Ability to detect an attribute in any position of the

image by sliding window.
 Tolerance to slight variations in rotation,
translation or scale, thanks to the pooling layers. Fig.3: Sample images dataset Chars74
 Ability to recognize complex attributes or
characteristics of high level by combining the However, not all images in this collection are valid for the
filters of the convolution layers. network training: lowercase letters and copies generated by
"exotic" sources they are not representative of the problem.
An example of the predominance of convolution networks Excluding these images collection generated characters,
is the results of ImageNet the prestigious testing [1]. In these computer is used in training to reduce to a total of 34,584
annual tests they complete teams of experts in artificial images of the digits 0-9 and letters AZ, mean about 950
vision of the whole planet presenting systems that attempt to images for each class of numbers and letters.
classify images into 200 predefined categories, starting from
a training set consisting of more than 450,000 images. Since 3.2 Data set to validate the system
the year 2012 to date, the systems based on CNN networks Get a dataset with actual images of vehicle license plates is a
that obtain the best results, with error rates considerably more complicated task than it might seem, since in many
lower than those offered by other artificial intelligence laws this information is considered a character data personal.
techniques[2]. In the case of Spain, a legal report of the Spanish Agency
Data Protection (AEPD) [4] concluded “the provisions of the
2. PROPOSED SYSTEM Organic Law 15/1999 on Protection of Personal Data (Act)
The system input images are of vehicle registrations. In this ". To avoid this problem has chosen to validate the ALPR
paper, we used both real images such as computer generated, system with a set of images of vehicle license plates mostly
automatically scaled to a resolution approximately 230x50 generated by computer, also including some real photograph
pixels for easy processing. Once an image is entered, the taken with vehicle owner's consent. These "synthetic"

49
 International Journal of Computer Applications (0975 – 8887)
Volume 177 – No.3, November 2017

license plates have been obtained using tools Free Web sites with vectors and matrices, fundamental when working with
platesmania.com [5] and acme.com [6]. images and artificial neural networks.

The first allows you to create registrations under existing 4.1.2 OpenCV
national format, two groups of characters consisting of a OpenCV is a library of artificial vision developed by Intel
four-digit number, from the 0000 to 9999, and three letters, and now released under BSD license, which has more than
starting with the BBB letters and ending with the ZZZ five hundred algorithms optimized to perform the main tasks
letters, where the five vowels and the letters Ñ, Q, CH and of computer vision, such as image processing, feature
LL are deleted. Furthermore, the tool website acme.com can detection or recognition objects. The library also has
generate state license plates according to different formats different learning algorithms machine as support vector
United States, but without any restrictions when entering machines (SVM), Naïve Bayes or KNN among others. It is
combinations of letters and numbers. written in C ++, multiplatform and has interfaces to work
with languages like Java or Python. The large number of
available algorithms, its execution speed and extensive user
community that have, they make OpenCV an indispensable
tool for developing systems artificial vision based on open
source software.

4.1.3 Tensor Flow

It is an open source library developed by Google to facilitate
design, construction and training systems deep learning.
Tensor Flow is characterized by a programming model in
which both the operation as data is stored internally in a
structure graph, which facilitates the display of dependencies
between operations and allocation to different devices such
as graphics processors. While it is true that there are other
libraries Deep Learning support Python that offer similar
Fig.4: Generator license plates acme.com
programming model; TensorFlow presents a fairly clear
syntax and has a powerful tool display ( Tensor Board ) that
Here are some examples of "synthetic" license plates are
allows us to understand, debug and optimize the graph
displayed generated with these tools to validate the system:
computing offering all kinds of statistics. In addition, Tensor
Flow behind the research team is project Google Brain,
which has some of the leading experts in Deep Learning as
Geoffrey Hinton and Alex Krizhevsky, authors of the first it
based on a CNN network that won the test in 2012 system
ImageNet [7].

4.2 IMAGE PROCESSING

As it was said in the system description, we first process
running once introduced the image of a license plate It is the
Fig.5: Examples of computer-generated license plates
character segmentation, which aims to detect those regions
image containing numbers and letters that subsequently try
4. TECHNOLOGIES EMPLOYED to identify the neural network classifier. The Python module
4.1 Tools that executes this process is based on OpenCV and employs
The main tools used to develop the ALPR system have been the following image processing techniques:
the Python programming language and libraries Tensor Flow
and OpenCV used to generate the convolution network and 4.2.1 Conversion to grayscale
system pre image processing of enrollments respectively. The color information is not required for the problem posed
and elimination facilitates the detection of the characters of
4.1.1 Python the license plate. for this therefore, the first transformation is
Python is an interpreted programming language for general applied to the image is a conversion to grayscale, or what is
purpose, multiplatform and multi-paradigm (imperative, the same, convert the image into a matrix in which each
object-oriented and functional). It is a strongly typed value represents the gray level of the corresponding pixel.
language, i.e. you can not assign one varying a value of a This transformation is performed by the function cvtColor of
different type to the original without an explicit conversion, OpenCV.
being turn this typing dynamic: the type of the variables is
determined in time execution depending on the assigned 4.2.2 Gaussian Blur
value. The reasons for the choice of this language are a clear The second transformation is done is to apply an image
and intuitive syntax. It facilitates code readability, and the effect smoothed to remove noise. Specifically, a blur filter
large number of specialized bookstores available, such as applied Gaussian mix consisting slightly gray levels between
Numpy, which allows a simple way all kinds of operations adjacent pixels, which results in an image with edges more

50
 International Journal of Computer Applications (0975 – 8887)
Volume 177 – No.3, November 2017

soft where small details are lost, similarly to what it occurs parameters to the needs of the problem raised in this paper.
in blurred pictures. This transformation is performed by This design also follows the general strategy proposed by
function Gaussian Blur of OpenCV. Simard et al [9] for visual analysis of documents by
convolution networks, it consist extract simple features of
4.2.3 Thresholding the characters in the first network layers, and later turns
The thresholding or "threshold technique" is a method that them into complex features thanks to the combination of the
allows you to converting image in black and white setting a various filters of successive layers convolution.
value from which all excess pixels that are transformed into
a binary color (white or black), and rest counter. In the case 5. IMPLEMENTATION
of this vehicle registrations transformation creates images 5.1 Data processing
which clearly fall defined outlines of the characters,
Then the various operations are described preprocessing they
facilitating the process of segmentation or isolation areas
carried out in the system, both the segmentation module
containing them.
characters, such as during the loading process image set
training the neural network.
In the proposed system this transformation is performed by
the function adaptiveThreshold OpenCV, wherein the
threshold value is calculated for different regions of the 5.1.1 Pre-processing input images
image, providing good results even if there illumination The various processing techniques described images that
variation image. apply to enrollments coming into the system - conversion
grayscale, Gaussian blur, thresholding and detection
contours - in order to isolate those regions that may contain
4.2.4 Edge Detection or letter of enrollment. Once identified these "regions of
The contours or edges are sets of curves joining adjacent
interest" (ROI), the Python module segmentation process
pixels having the same color or intensity. These curves can
manager performs a series of checks to determine if their
locate the borders of the objects in the image, and its
content is effectively a digit or a letter that is to be analyzed
detection is essential for artificial vision system can
and classified by the neural network. These checks are
recognize or detect shapes in a image. In the proposed
basically checked whether the height, width and pixel aspect
system this detection is the last stage of pre-image
ratio i.e. the ratio between width and height the region of
processing of tuition, and is carried out by the functions
interest - are within certain values. The minimum and
findContours and boundingRect of OpenCV.
maximum values of these parameters are determined by the
proportions they should have the characteristics of a standard
4.3 NEURAL NETWORK enrollment approximately 230x50 pixels, being the
ARCHITECTURE resolution to which you scale all the images that enter the
The main component of ALPR system presented in this system. If the proportions of a given region are valid, i.e.
work is its convolution neural network (CNN). This is they are within limits, the next step is to extract its contents
responsible for classifying characters of the license plates and save it as an image of 32x32 pixels that can be analyzed
from images taken in the module segmentation, for which a by the CNN network system.
prior process of training is necessary for the network to learn
the characteristics that define the different letters and 5.1.2 Pre-processing training data set
numbers that may be present in an enrollment. To train the CNN network is necessary to import the subset
of representative images of the problem that contains the
dataset Chars74K in a data matrix. Each subset of the
dataset is composed of images in grayscale computer
generated characters corresponding to the numbers 0-9 and
uppercase letters AZ. The big advantage of using a network
neuronal as ranking algorithm is their learning ability
automatic, thanks to which it is not necessary to develop
extraction processes attribute or use techniques to reduce the
dimensionality as PCA ( Principal Component Analysis ) or
the like; greatly simplifying the pre-processing tasks of the
training data.

5.2 Network training

The training process of the CNN network system is based on
the combination algorithm back propagation with the drop
Fig.6: Architecture of the convolution neural network method gradient. It is basically an optimization problem in
proposal which iteratively weights and - the values of the parameters
are searched bias network – making the value of the error
This architecture is based on the convolution network LeNet- function or cost is minimal. For network CNN system, the
5 proposal by Yann Lecun [8] for recognition of digits, selected function to measure how good or bad the
adapting the number of layers and values of various classification of images is the error of cross entropy (cross

51
 International Journal of Computer Applications (0975 – 8887)
Volume 177 – No.3, November 2017

entropy Error), one of the usual design of convolution neural

network. Of the different variants of gradient descent
method has been chosen version stochastic or incremental ,
that is to calculate iteratively the gradient for a subset or
batch of training data, updating the network weights by
propagating errors to back, without waiting for the full data
set is processed. From the standpoint of computational cost,
this method is more efficient to process the gradient in each
iteration the dataset completely, because usually provide
faster convergence. Although it is also true that runs the risk
of being "stuck" in a local minimum of the error function.

5.2.1 Initialization
Before running the training process is necessary to generate
and initializes by Tensor Flow network architecture. Briefly,
in the code module, A declares graph operations that
contains both the variables that store data training, weight
matrices, bias and outputs of the network; as calls the Fig.7: Evolution of precision during training
primitives framework to define the error function, the layers
convolution (with window stride, filters and activation As you can see the distance between them is very small,
functions), layers pooling, layer FC, exit softmax and the call which a priori is a good indicator that has not been produced
to update weights by back propagation and gradient descent. over fitting. You can also observe the high accuracy obtained
In the next picture you can see the graph of the CNN in both subsets: about 99% to 98% and training for
network system exactly it represents as the display tool validation. For the set of images of test accuracy obtained
Tensor Board included in the framework. Besides declare with the settings defined in the preceding paragraph is
the structure of the network, in operation network indicated 97.83%. Regarding the role of errors, the following chart
also how it should be initialized. shows its 4,500 iterations. Here you can see a fall very
pronounced during the first 500 iterations and quite softer
5.2.2 Training but constant for the rest, indicating that in principle the
Once declared and initialized the network, the next step is to learning rate chosen for the gradient method seems
import the set of training images preprocessed by the code appropriate.
module. Before starting the training, this set of images is
divided randomly into three subsets called training,
validation and test: the first contains a total of 28,012
images, the second3113 and third 3459. The training set is
used to determine the weights of the network allow
classification with the lowest possible error images,
validation to check if over fitting (occurs over fitting ) during
training and test set to estimate the accuracy of the system to
classifying images that have not been analyzed during the
iterations of training. The section of code responsible for
implementing the training process - makes a total of 4,500
iterations selecting each one lot or batch of 300 images of
the training set to optimize network weights by function
Gradient Descent Optimizer of Tensor Flow with a learning
rate of 0.0001; which means that the end of the will process
each image is analyzed about 50 times approximately.

5.3 Evaluation
To evaluate the effectiveness and efficiency of the learning Fig.8: Evolution error function during training
process of the network has chosen to monitor the evolution
of the overall accuracy and error in classifying images of In addition to tracking the above figures, the assessment
subsets of training and validation during the iterative effectiveness of the learning process is completed running
training process. Tracking accuracy detect if overload is the network trained with a sample of digits and letters
occurring training (over fitting) network, i.e., if the weights extracted license plate images synthetic and real.
and bias are excessively adjusting to the peculiarities of
images. Training subset, which may cause once trained the
network is not able to correctly classify the different images 5.4 OPTIMIZATION
analyzed during learning. In the next picture you can see the The characteristics of the network architecture (layers,
evolution of precision in both the training set and the filters, window size, etc.), as learning process parameters
validation for 4500 iterations. (number of iterations, batch size, learning rate, etc.); they

52
 International Journal of Computer Applications (0975 – 8887)
Volume 177 – No.3, November 2017

have been determined empirically in particularly following 7. CONCLUSIONS

the procedure proposed by Nikhil Buduma [10]. It is The first conclusion is the confirmation of the feasibility of
adjusted iteratively parameters network architecture and automatic license plate recognition (ALPR) based on a
training process function results. Apart from the above convolution neural network. The use of these networks
adjustment process diagram for optimization network have creates a system really capable of "learning" the
also followed some of the recommendations and techniques characteristics of the different digits and letters of the license
common in the design of convolution neural network [11], plate, without complex mechanisms attribute extraction. The
namely: second is the realization of the importance of having a set of
• Review training data set to delete images not images of sufficiently extensive training and representative
representative of the problem. Here is the representations of the problem to obtain satisfactory results with a
exotic or ambiguous character. To detect them, apart from convolution neural network. In the case of the system
manually review the dataset, it has been of great help assess presented in this work, the dataset used was not specific to
the mistakes made by the network with sample images the problem, and proof of this is the large number of images
enrollment during training. of characters in lowercase or exotic sources that had to be
eliminated to optimize the learning process. The third
• Work in learning algorithms machine consists of conclusion is so laborious that can be optimized learning
subtracting the average data matrices each column. This process convolution neural network, and poor about
focus usually comes accompanied by a normalization of data literature there currently. During the development of this
[13-16], in the case of images is not necessary since the paper has been necessary to repeat many times the proposed
values of pixels are on the same scale (0-255). process to give architecture and process adjustments training
• Dropout of neurons fully connected layer (FC) is about offered reasonably satisfactory results.
fairly new technique to reduce overtraining. It was by
Srivastava et al [12] in 2014 and it is "off" randomly some We can extend this work by adding a pre-segmentation
neurons during training, so only the weights of the remaining module to allow characters automatically locate vehicle
active are updated. registrations in photographs. Review the segmentation
module to avoid the use of measures absolute in pixels to
detect regions of interest. Expand the dataset training,
6. RESULTS OBTAINED gathering new images or by using techniques such as “data
Once trained network CNN is just check the effectiveness of augmentation “to create new images from existing
license plate recognition system designed. To this end they transformations, which undoubtedly improve the percentage
have been used three collections of images, two of them of correct system.
created with tools. Details of each are as follows:
1. Collection of 100 images of randomly generated by
licensed plates computer according to the current national
8. REFERENCES
format, i.e. two groups characters consisting of a four-digit [1] ImageNet Project. http://image-net.org - March 2017.
number, from 0000 to 9999, and three letters, where the five [2] ImageNet ILSVRC2015.http://image-
vowels and the letters Ñ, Q are deleted, CH and LL. net.org/challenges/LSVRC/2015/results - March 2017.
2. Collection of 100 images of randomly generated by
licensed plates US computer according to different formats. [3] Chars74k dataset. TE Campos, BR Babu, M. Varma.
Since there is no a single standard in this country, has chosen http://www.ee.surrey.ac.uk/CVSSP/demos/chars74k/ -
to use one of the common formats, consisting of two groups March 2017.
of three characters, the first with numbers 000 to 999, and [4] Report 425/2006. Spanish Data Protection Agency
the second with letters from AAA to ZZZ. (AGPD).
The results obtained for the three collections are: [5] Platesmania License Plate Generator.
http://platesmania.com/es/informer - April 2017
Total Correct Failures [6] Acme License Maker. http://acme.com/licensemaker -
Collection images hit hit April 2017
National Matriculas 100 100 0 [7] ImageNet Classification With Deep Convolutional
Neural Networks. Krizhevsky et
Matriculas USA 100 60 40 al.http://papers.nips.cc/paper/4824-imagenet
classification- with-deep-convolutional-neural-
networks.pdf March 2017
The results for the 100 images in national format are really [8] Gradient-Based Learning Applied to Document
spectacular, obtaining 100% of correct recognition content Recognition. LeCun et
of these images of license plates. In the case of real images al.http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf
the results are equally impressive 100% hit although it is - March 2017
important to note that this sample is too small to be
considered significant and also images they have been [9] Best Practices for Convolutional Neural Networks
obtained on highly concessional terms. Applied to Visual Document Analysis.Simardetal.

53
 International Journal of Computer Applications (0975 – 8887)
Volume 177 – No.3, November 2017

http://research.microsoft.com/pubs/68920/icdar03.pdf - Transfer from a Permeable Horizontal Isothermal

April 2017 Cylinder." Frontiers in Heat and Mass Transfer
FHMT) 9, no. 1 (2017).
[10] Fundamentals of Deep Learning. Nikhil Buduma.
O'Reilly. [14] Rajendra, P., et al. "Potential Distribution in a Nano
http://shop.oreilly.com/product/0636920039709.do - Channel of an Electrolytic Solution by Finite Element
April 2017 Method." I J C T A, 9(32), 2016, pp. 145-151
[11] Deep Learning Book. Ian Goodfellow, Yoshua [15] Padidhapu, Rajendra, Shahnaz Bathul, and V.
Bengio,Aaron Courville. MIT Brahmajirao. "Least Square and Gauss Jordan Methods
Press.http://www.deeplearningbook.org - March 2017 applied to fit the Dielectric data Vs. Concentration of
Ionic Liquids." International Journal of Science and
[12] Dropout: A Simple Way to Prevent Neural Networks Technology 2.1 (2013).
from Over fitting. Srivastava et al.
https://www.cs.toronto.edu/~hinton/absps/JMLRdropou [16] Boadh, Rahul, et al. "numerical simulation of boundary
t.pdf - May 2017 layer flow parameters by using wrf-arw model over a
tropical region.", Jr. of Industrial Pollution Control
[13] Subba Rao, A et al. "Numerical Study of Non- 33(1) (2017) pp 1148-1154.
Newtonian Polymeric Boundary Layer Flow And Heat

IJCATM : www.ijcaonline.org 54