Final ThesisV1

I'm here to present to my final thesis which is

based on a comparative study to classify the denomination and

detection counterfeit of indian magnets.

So with no further delay, let's start with the presentation.

I would remove my uh huh better front background and

you have to move into the presentation.

So my presentation consists of seven slides, seven sections.

Introduction, problem statement aims and objectives, literature, review, research methodology,

analysis and discussion, conclusion and recommendation.

Yeah. Let's move ahead.

So an introduction.

The moment organizations plays a crucial part.

Gas transaction can easily be a victim of their currency.

Annual report in 2018 says that in India it's a

big concern fake news of a big concern.

There has been around 5 22,000 and 8 to 300

ft notes which are available in the market.

We need to stop this and to stop illegal business

activities and economical terrorism.

What is if I see and so if I see

any is fake indian currency note, which is usually used

by the National Crime Records Bureau in India and the

media to address fake notes.

Uh, so how can machine learning help us uh to

identify fake notes?

So advancement in machine learning has been created.

Uh, the image recognition especially they're really cost effective, accurate

systems and portable and easy to use.

Uh, so we're here, we're also planning or aiming to

create a cost effective and efficient way of identifying.

Banknotes are using deep learning and image recognition techniques related

works. So most of the related works were based on

optical or electrical sensors, uh focusing on the inbuilt photos

instead of holographic features of the notes are based on

images or visual attributes which are classified to accept features.

This is mostly used in deep learning and machine learning

techniques. Let's go through some examples of previous papers or

study by military dancing, which was focused on Canaan for

classification and S.

V. M.

Of a classification and H.

O. G.

Was used for image feature extraction, which was around 90%

accurate. Then there was another study where serena nadine in

which a hybrid model was used using ocr simple face

recognition model and half transformation for ultraviolet clients detection.

It was around 93% effective.

Then there was another study by roger in which metal

and Canaan techniques were used.

Uh These images were cropped segmented and the intensity of

the features are being used for the detection of fake

notes. What are the problems with these studies?

The limitations that were generally noticed were uh they were

really gen rick uh for all the denominations uh limited

studies on more updated and efficiency nN models uh Color

spaces were not really experimented with.

All invalid notes were not considered because they can also

be part of fake notes are a problem for visually

impaired. What makes this tastes different?

So for these species, we have considered all the denominations.

Uh Different color spaces have also been taken into account

both handcrafted and techniques.

Both hand capture techniques and CNN techniques were used uh

for classification.

What is the aim and objective to create a system

that is accurate uh able to successfully classify denominations and

identify fake notes.

Outline of the objectives.

To use a traditional and modern machine learning techniques to

use hand crafted and also deploying techniques to extract features

to do a comparative study of the models they can

in SPM and CNN uh and then suggest the most

appropriate and optimized approach.

Uh significance and scope, contribution to successful bank no detection,

uh including comparative study, which can be later used uh

for further improvement Literature review in 2016, New currencies were

introduced and the whole currencies, all denominations of 500 and

100 Anna were not legal anymore.

So modern image recognition can make use of this newly

visually different banknotes introduced for fake node dissection.

Let's discuss some of the security features in the new

banknotes security thread, there's a watermark micro lettering latin image

portrait of Mahatma Gandhi which are also there previously.

Then they're bleed lines and then see to register optically

variable ink and a panel guarantee close Here are two

images of Iron Ore 500 showing the features of the

new banknotes.

All right.

So how image recognition is done using neural networks.

So neuron, it was basically replicates or try to replicate

human brain, how it works or is being inspired by

it. Uh So it has been used to successfully to

attain a lot of complex activities like speech recognition image

processing. There are also successful implementation in the healthcare industry

for complex structure analysis has been also used to develop

intelligent traffic system and there are a lot of other

examples like this.

So what are the techniques that are being used here?

We have used a Canon model can and model is

to kenya's algorithm which is basically focused on the distance

of the data points and classify the images based on

the features which are closest to it.

It's it's a very, quite popular um algorithm for classification.

Then there is S.

V. M.

S. V.

M belongs to linear machine learning models.

They are extremely efficient, extremely accurate, Especially good when there

are only two groups for classification, then there is of

course the neural networks, uh, neural networks are extremely advanced,

they're really efficient.

Uh They have this tendency to improve their performance with

more and more training.

And it is, it's so it's quite advanced than usually,

uh more advanced in the traditional machine learning algorithms.

Uh, so for testing uh, train split, we are used

to cross validation, especially the careful cross validation over here.

Cross validation help in eliminating under fitting and overfitting problems.

Um Kay fooled.

Cross validation is especially helpful when the number of data

sets are really limited.

Mm And there is also a ah class imbalance in

which case we use stratification or stratified, careful cross validation.

Let's move into research methodology.

So over here, the system is that working on images

of indian magnets for verification and classification.

Uh, We have considered both front and rear view of

the banknotes.

So, and we have included both modern and all magnets

that are currently in use.

So let's look at the workflow utilization of button both

front and back view of images.

Uh, Traditional machine learning algorithms has mused again.

NNS VM.

Also CNN uh neural networking has been used.

H. I.

G. Has been used for uh feature extraction in hand

grocery techniques.

Are we able to use the modern C.

And techniques for feature extraction?

Um Kay fold, stratified, careful cross validation has been used

for testing and split a sudden evaluation metrics have been

taken into consideration like area, under curve classification, accuracy, scoring

position. School different colours, species had also been accounted for.

This is basically the workflow.

The data moves into the pre processing state.

Uh It's processed and feature extraction is done.

Classification model classified into different denomination, invest on the denomination

denomination. It is uh fetch to that particular flow where

the features are getting extracted.

We referred to that classification model and then authentication is

done. The data sets includes both self taken and images

which are available online.

Uh It included different images with different backgrounds and different

lighting conditions.

Pre processing was stunned where it was passed through a

caution filter to remove noise and smoothing was done.

Then they were transformed into two separate color spaces.

RGB NLB uh to compare the performances.

Feature extraction was done during one of the handcrafted techniques

of H O.

G. And then a neural net was like VG 19

and squeezed.

It train interests split was done using stratified, careful cross

validation technique as discussed.

Earlier evolution metrics included a UCC F on precision and

recall a from school and recalled school were given the

highest priority order based on the requirement.

Let's look into the work floor so the images through

data preparation, flu then data pre processing.

The middle was transformed.

Feature was extracted, shouldn't desperate was done.

And then it was fed to the classification models.

All the NATO was classified.

And then we did the performance evaluation.

Uh huh.

So data filtering was done using Goshen filter remove noise.

Is this is a picture of a lab color transformed

image of the note.

And then we moved into feature extraction in the first

phase, the north for classified into different denominations.

Uh and then uh feature was extracted and then fed

to a different classifies of KMs human.

And um ah then in the second phase After the

denomination was considered uh then it features were extracted using

squeeze net and GT 19 and Hawk feature.

Um The features are extracted for both RGB NLB transformed

images and then for two networks classifies like ken and

feminine. And uh let's look at the training workflows.

Both kind of transformed images like RGB NLB affair to

feature extraction, which is done using hog 19, squeezing it

and then train desperate was done and then fit models

again and spm and neural networks and then model liberation

was done on them Discussion.

Okay, so for classification of notes and it was performed

the best with an F one score of 0.907.

Uh For classification of, not based on authenticity is being

performed the best.

Um When the model was trained on RGB images, squids

net uh was performing better for seeing model extraction and

it was not the best in the outcome, but the

performance was consistent.

CNN uh Squeezing it improved the performance because it was

faster and more efficient and which is 19 and CNN

techniques took longer than handcrafted techniques.

Uh but they were much more accurate and performed better

than H O G mhm extraction technique.

and it takes a longer time level.

Of course within.

It needs more consistent and takes a fraction of time.

More. Rgb color space proved to be more effective.

Which was really surprising because early because of space to

considered very close to humanize perception uh turned out to

be less effective.

So conclusion recommend recommended to use quiz net for image

and merging and invest.

Model can be used to classify the north based on

denominations. After determining the denominations.

Those can be fed to an SV model to detect

its authenticity.

For emails, denomination is the model should be trained for

detection of technique.

Notes specific is via model increases the accuracy and effectiveness

of the model dramatically specific when it is specific to

denominations rather than generate models.

Uh So what are the future recommendation in future?

A system?

We recommend to create a system that can detect multiple

notes of different denominations in a single scene and calculate

the total valuation.

That will be really good and helpful in real life.

Uh Then we can also enhance the present study through

various augmentation techniques applied on the images, like for example,

cropping them and the object and eviction cropping it and

changing its orientation.

And then individual feature detection counters for security features to

check authenticity of the notes.

Um This can prove to be much more efficient and

much more accurate.

Ah So that's it from a presentation.

Uh It was a pleasure.

Thank you.

Have a nice day.

Final ThesisV1

Chapters

Audio TranscriptAuto-generated