As Andrew NG phrases it perfectly “AI is the new electricity”. Signzy is true believers of this and have witnessed what AI could do.Just watch the PPT you will come to know how signzy replaced legacy banking processes with AI-driven technology.
2. Introductio
nAs Andrew NG phrases it perfectly “AI is the new electricity”. We at Signzy are
true believers of this and have witnessed what AI could do, how deep its impact
can be and the drift it has brought in the way we approach our problems. Being in
the finance and regulatory domain we have been identifying a series of our
customer problems which could potentially be addressed through AI.
In this PPT we will take you to behind-the-scenes of our state-of-the-art system
and how we tackled the problem, ultimately overpassing the targeted accuracy
required for real world use.
3. Knowing the beast we are
to fightThe idea is to deploy the pipeline into financial institutions with all possibilities
of input variation and yet it should surpass or at least be equivalent to accuracy
of a human being. The solution is to work on data which arrives from the most
rural parts with pics taken from even 0.3 MegaPixel cameras and travelling over
a dramatically slow connectivity. We knew the toughest challenge was to
cater to variations that could arrive in inputs.
But at last We were able to get desired training sets successfully, which was
half the problem solved.
4. • Our target was to create the architecture which could be more than 99%
accurate and yet be fast enough to make an impact.
• You can’t reject an application by an old rural lady, who has brought you a
photocopy of printout which in turn is obtained from a scanned copy of a
long faded PAN card. We took it as a challenge to create the system so that
it can help even the rural Indian masses.
World is not the cozy
laboratory, we know that!
5. Creating the technology
Baby steps ahead
We tried out various approaches for solving the problem. Firstly we extracted
features using Histogram of Oriented Gradients (HOG) feature extractor from
OpenCV and then trained a Support Vector Machine (SVM) classifier on top of
the extracted features. The results were further improved by choosing XGBoost
classifier. We were able to reach about 72% accuracy. We were using Scikit
learn machine learning framework for this.
6. Not enough, let’s try
something elseIn our second approach, we tried ‘Bag of words’ model where we had built a
corpus containing unique words from each identity card. Then we feed the
test identity cards to an inhouse developed OCR pipeline to extract text from
the identity card. Finally we input the extracted text to a ‘Naive bayes’
classifier for the predictions. This method boosted the accuracy to 96% . But
the drawback of this approach was that it can be easily fooled by hand
written text.
*Many more approaches were also followed*
7. The final approach is where the novelty of our algorithm lies. The idea is to use
an image object detector ensemble model for image classification purpose. For
eg. the Aadhaar identity has Indian Emblem, QR code objects in it.
Hit the right nail kid, treat
them as objects
8. Good work lad!
What next?• While the pipeline we had come up with till here has a very high
accuracy and efficient processing time, it was yet far from the a
productionised software.
• We are clearly seeing the impact deep learning can have on solving these
problems which we once were unable to comprehend through
technology. We were able to gauge the huge margin of enhancement that
deep learning provides over traditional image processing algorithms. It’s
truly the new technological wave. And that’s for good.
9. Contact Details
Bangalore office
BPL building, 11th KM, Arakere
Bannerghatta Rd,
Bengaluru, Karnataka – 560076
Mumbai office
“Rise” 1902, 19th Floor, Peninsula
Business Park, Tower B, Lower Parel,
Mumbai – 400013
Website : https://signzy.com/Email : reachout@signzy.com
About Signzy:
Signzy helps financial institutions transform current semi-manual processes
into real-time digital systems, using Artificial Intelligence and Blockchain. This
ensures that the new processes are user-friendly, yet secure and compliant.