0% found this document useful (0 votes)
39 views5 pages

Irjet V8i5131

Uploaded by

Akanksha
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
39 views5 pages

Irjet V8i5131

Uploaded by

Akanksha
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 08 Issue: 05 | May 2021 www.irjet.net p-ISSN: 2395-0072

Handwriting Forgery Detection using NN

Siya Philip1, Shikha S Nambiar2, Shreya J3 , T V N Satya Pratyusha4 and Sneha S Bagalkot5
1Student, Dept. of Computer Science Engineering, Presidency University Bangalore, Karnataka, India
2 Student, Dept. of Computer Science Engineering, Presidency University Bangalore, Karnataka, India
3Student, Dept. of Computer Science Engineering, Presidency University Bangalore, Karnataka, India
4Student, Dept. of Computer Science Engineering, Presidency University Bangalore, Karnataka, India
5Assistant Professor, Dept. of Computer Science Engineering, Presidency University, Bangalore, Karnataka, India

---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract – Handwriting is unique to each person, much Specific features of Handwriting are:
like a fingerprint. Since every handwriting is unique, it is I. The roundness of the letters
also referred to as the brain's fingerprint. Criminals use II. Spacing between letters
handwriting forgery to fraudulently produce, change, or III. The pressure put on paper while writing
write a person's handwriting such that it appears similar to IV. The average size of letters
the real handwriting in most cases, with the intent of V. The inclined angle of letters
profiting from the innocent party. In this present study, a
method has been proposed where the model is trained with The above are some characteristics that aid in determining
a dataset of handwriting, and predictions are made as to the authenticity of a person's handwriting. Since they will
whether a provided signature is genuine or forged based on be unique to each person, these characteristics can be
the features like ratio, centroid, eccentricity, skew and used to make decisions.
kurtosis, and solidity of the words.
We tried to assess the validity of a given input by
Key Words: Handwriting Forgery Detection, Word considering variables and comparing them in this project
Segmentation, Image Pre-processing, Feature Extraction, because this is an essential consideration when
Multi-Layer Perceptron, Neural Network, Prediction. considering the trueness of the text.

This paper is our take on reducing the chances of forgery


1.INTRODUCTION by comparing the handwriting of multiple users and
training the model to recognize the real and forged inputs.
Handwriting is often used to assess a person's personality.
It is very normal for certain experts to be able to predict a
2. PREVIOUS WORK
person's behavior based solely on their handwriting. In a
similar vein, we're attempting to develop a framework
Using previous work on the subject as a source of
that calculates some characteristics to determine the
inspiration [8] In this research paper, they proposed a
validity of a note.
text-independent handwriting forgery detection system
based on branchlet features and GMMs. They divide the
The internet has altered how the environment operates.
function data into groups and use an exhaustive method to
During this transition, we tend to exchange a lot of
create their own GMMs. They then compare each
documents over the internet because we are still
category's similarity to the other input data.
transitioning to a completely online mode where paper is
seldom used.
3. PROPOSED SYSTEM
Since we are sending a physically printed document that
does not contain any digital signatures that can be 3.1 Word Segmentation
checked, the documents can easily be forged during this
process. In this way, we put ourselves in a vulnerable We will use Scale Space Technique for Word Segmentation
position, allowing the forger to easily alter the text. This to segment words from a given input image [7] where a
could result in a financial loss or unauthorized changes to grey level image is used as the system's input. The image is
a legal document. That is the reason Forgery is considered processed to eliminate horizontal and vertical line
a white-collar crime. segments that could cause problems during subsequent
operations.
To avoid such situations, we should ensure the document's The page is then dissected into lines using grayscale image
originality, which we can do by conducting a Handwriting projection analysis methods that have been updated. The
forgery search. This will assist us in being safer. projection function is smoothed with a Gaussian filter (low

© 2021, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 634
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 08 Issue: 05 | May 2021 www.irjet.net p-ISSN: 2395-0072

pass filtering) to eliminate false alarms, and the locations the as feature extraction will now be able to give more
of the local maxima (white space between the lines) are accurate values.
determined. Line segmentation is useful for breaking up Second, we then remove any noise disturbances using
connected ascenders and descenders as well as deriving Gaussian filter in the image and then convert it into the
an automatic scale selection mechanism. binary format as it will be very easy for us to consider and
compare them later on.
To build a scale space, the line images are smoothed and
then convolved with second-order anisotropic Gaussian
derivative filters, and the blob-like features that provide
us with the focus of attention regions (i.e., words in the
original document image). A connected component
analysis of the blob image is used to extract words, which
is accompanied by a reverse mapping of the bounding
boxes. After that, the box is vertically extended to make Fig -4: Genuine Image Fig -5: Forged Image
room for the ascenders and descenders.
Images after Pre-Processing:

Fig -1: Genuine Text

Fig -6: Genuine Image Fig -7: Forged Image

Fig -2: Forged Text 3.3 Feature Extraction

This process identifies important features of the image


and gets their values and stores them in one place. Then
we compare these values among the images to get the end
result. Well, there are numerous features that can be
considered we have listed out a few which according to us
will give accurate values. This particular step defines the
efficiency of the software.
The following are the features whose values will be
extracted from the image [3]:

I. Ratio: This calculates the relationship between


the height and width of the image. It basically describes
the shape of the image.

ii. Centroid: The centroid of the image is often


considered the intersection point of all the hyperplanes of
Fig -3: Word Segmentation and Dataset Formation symmetry within the image, by doing this we get the
center point of the image.
3.2 Pre-Processing
iii. Eccentricity: The eccentricity of an ellipse is
Pre-Processing is an essential step to perform before we the ratio of the distance between its foci to the length of its
dive into the comparisons as this helps in enhancing the main axis. The value is always in the range of 0 to 1.
feature of the images and increases the efficiency of the
system. To increase the clarity of the image we have iv. Skew: When transferring data to a digital
followed a set of steps which are as follows: format, skew detection is one of the first operations
First, we convert the given input image into a greyscale performed on scanned documents. Its aim is to align an
format so that we now have the images only in black, image before it is processed because text segmentation
white, and shades of grey colors which will provide us the and recognition methods depend on correctly aligned
raw image. This can be helpful to increase the efficiency of next lines.

© 2021, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 635
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 08 Issue: 05 | May 2021 www.irjet.net p-ISSN: 2395-0072

take weighted input signals and use an activation function


to generate an output signal.[5]

Fig -8: Skewness


Fig -10: Weights and Activation
kurtosis: Kurtosis is a statistical indicator of how
often the tails of distribution vary from the tails of the
regular distribution. It assesses the sharpness of a a) Weights
frequency distribution curve.
Weights are often set to small random values, such as
those in the range of 0 to 0.3.

b) Activation

The weighted inputs are added together and passed


through a transfer function, which is an activation
function. It controls the frequency of the output signal as
well as the threshold at which the neurons are activated.

3.4.2 Networks of neurons


Fig -9: Kurtosis
There are several layers in a network.
VI: Solidity: The area of an image object is divided
by the area of its bounding rectangle to determine its
extent. The area of an image object divided by the area of
its convex hull determines its solidity. A fraction of the size
of your actual picture.

3.4 MODEL

We have built a neural network (multi-layer perceptron)


using TensorFlow and successfully train it to recognize if
handwriting is genuine or forged.

3.4.1 Multi-Layer Perceptron Fig -11: Layers of Neuron Network

Multi-Layer Perceptron is the most complex Artificial a) Input/Visible Layer


Neural Network architecture, fundamentally made up of
several layers of the perceptron. For supervised learning, Since it is the exposed component of the network, the
MLP networks are used. It's a feedforward Artificial bottom layer that takes input from the dataset is called the
Neural Network that takes a set of inputs and produces a visible layer. Typically, a neural network is drawn with
set of outputs. An MLP has several layers of input nodes one neuron per input value or column in the dataset and
that are linked as a directed graph between the input and one visible layer.
output layers. The ability of a Neural Network to learn the
representation of training data and how to relate it to the b) Hidden Layer
output variable that we need to predict is what gives it its
strength. The hierarchical or multi-layered structure of a The hidden layer is the next layer after the input layer.
Neural Network contributes to its predictive performance. They aren't exposed to the input directly. The most basic
network structure is a single neuron in the hidden layer
Artificial neurons are the building blocks of a neural that outputs the value directly.
network. These are essentially computational units that

© 2021, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 636
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 08 Issue: 05 | May 2021 www.irjet.net p-ISSN: 2395-0072

c) Output Layer value. The weight aids in the connection of one layer to the
next.
They are the final hidden layer, and they are in charge of
the problem-related output values needed for the problem The first layer's input value and weight are multiplied and
statement. added to the bias of the first layer, and the output is used
as the input for the next layer.
3.4.3 Training model

a) Data set

The first step is to read the train and test data. We begin
by reading the training CSV file and performing operations
on it, such as retrieving data from a column and storing it
in an array using call values. Fig -12: Weights and Bias

Then we use the astype() function, which creates a new This is how we develop our model, and we build it by
copy of the training input with each value converted to a passing the input value.
float (in our case). It doesn't change the training input, so To feed data into the tensor flow graph, we use a
you can check the value returned by astype() to get the placeholder.
converted array.
d) Loss and Optimizer
The method to categorical() can then be used to transform The next move is to locate the optimizer and loss. The aim
a NumPy array with data representing various categories of optimization is to reduce the loss function to the
into a NumPy array with binary values. It has the same smallest possible value. If the loss is reduced to an
number of rows as the input array and the same number appropriate amount, the model will learn an indirect
of columns as the number of classes. function that maps the input to the output.

On the testing range, a similar function is performed. The e) Accuracy


model receives this result of binary values as data. We'll now figure out how accurate each epoch is. The
accuracy of our model is the percentage of predictions that
The default graph stack is then cleared, and the global were right.[6]
default graph is reset.
f) Prediction
b) Neuron Layers The image prediction would then include an array of two
columns, the first column containing the score of
To work with the Tensor, we define a few main confidence against genuine and the second column
parameters and variables, such as learning rate, training containing the score of confidence against forged, based
epochs, and display steps. The number of hidden layers on the measurement accuracy. If the first column value
and neurons for each layer is then determined. There are exceeds the second column value, the document is
three layers in total: one input layer, two hidden layers, genuine; otherwise, it is forged.
and one output layer.
This model is used to determine whether or not a
handwritten document is authentic.
c) Weights and Bias
4. IMPLEMENTATION AND RESULTS
Weights and biases are assigned to each of these layers.
When inputs are transmitted between neurons, the Subjects were asked to write test samples in their natural
weights and bias are applied to the input. Weights handwriting style as well as forge handwriting samples
determine the strength of the interaction between from other subjects. Handwriting samples were scanned
neurons, or in other words, how much influence the input and saved digitally as a part of the process. The writing
has on the output. samples were then used to compute word-level
features.[4]
The previous layer has no impact on bias units, and they When we get the dataset from the user, we let each image
have no outgoing relations with their own weights. Before go through pre-processing. Once we get the raw images
moving the data to the next layer, a single node multiplies that would give us more accurate values, we perform
the input data by an allocated weight and adds a bias. feature extraction in which we extract values for ratio,
centroid, solidity, eccentricity, skewness, and kurtosis.
The two parameters vary in the degree to which they
In the given figure we have considered a dataset of 10
affect the input data, with Bias responsible for the
users taking their real handwriting and also made a 3rd
difference between the function output and the intended
© 2021, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 637
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 08 Issue: 05 | May 2021 www.irjet.net p-ISSN: 2395-0072

person write the same word which we will consider as the forgeries of the handwriting of other subjects. These
forged image. Each of these 10 users have given 5 images handwritings were scanned digitally and saved in a folder.
as in a user with id no 1 has given 5 samples of his real
handwriting and also 5 of the forged ones. This is so that As a result, we were able to distinguish between genuine
when performing comparison of values at the end we can and forged handwritten documents using pre-processing,
get an efficient result if we have more values. feature extraction, and training the model with genuine
Now we transfer these to function extraction, which and forged image datasets.
returns their values in CSV files. As can be seen in Figs 6
and 7, the device stores three of the values from five actual REFERENCES
images in the Training CSV file and the other two in the
Testing CSV file. [1] Navin Karanth , Vijay Desai and S. M. Kulkarni. 2011.
As can be seen in Figures 6 and 7, the situation for the 5 Development of an automated handwriting analysis
forged photos is identical, with three of them being sent to system
the Training file and the other two being sent to the
Testing file. [2] Amr Megahed, Sondos M Fadl Harbin and Qi
Han.2017. Handwriting forgery detection based on ink
colour features

[3] S JeromeGiden, AnuragKandulna , Aron Abhishek


Kujur , A diana and KumudhaRaimond.2018.
Handwritten Signature Forgery Detection using
Convolutional Neural Networks

[4] Sung-Hyuk Cha and Charles C. Tapper, Automatic


Detection of Handwriting Forgery
Fig -13: Training Dataset
[5] Agarwal, A., 2018. Multi-Layer perceptron using
Tensorflow. [online] Medium. Available at:
<https://towardsdatascience.com/multi-layer-
perceptron-using-tensorflow-9f3e218a4809>

[6] Deb, S., 2017. Neural Network Tutorial —Multi Layer


Fig -14: Testing Dataset Perceptron. [online] Medium. Available at:
<https://medium.com/edureka/neural-network-
To do the final verification on whether the handwriting is tutorial-2a46b22394c9>
genuine or forged, we ask the user to enter a specific
person's id for whom they want to check the handwriting. [7] Manmatha R., Srimal N. (1999) Scale Space Technique
for Word Segmentation in Handwritten Documents.
We take a user's input image, perform feature extraction In: Nielsen M., Johansen P., Olsen O.F., Weickert J. (eds)
on it, and save it as a new csv file. The model now Scale-Space Theories in Computer Vision. Scale-Space
compares these values to the testing and training csv files, 1999. Lecture Notes in Computer Science, vol 1682.
determining if the nearest value is genuine or forged. Springer, Berlin, Heidelberg.
https://doi.org/10.1007/3-540-48236-9_3
It then tests the output value of the closest value of
[8] Chin-Shyumg Fahn,Chu-PingLee and Heng-I Chen
features and shows it as a real image if that value is 1, and
.2016. A Text Independent Handwriting Forgery
as a forged image if that value is 0.
Detection System Based on Branchlet Features and
Gaussian Mixture Models
5. CONCLUSIONS

In this paper, we presented a handwriting forgery


detection system using a multi-layer perceptron. This
method assists in determining whether a handwritten
document is authentic or forged. The system accomplishes
this by measuring the precision and making an
appropriate prediction.

We collected experimental handwriting data from subjects


who wrote examples of their own handwriting and

© 2021, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 638

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy