How deep learning can help us in automation of document verification

I think I have spent 80% of this weekend uploading documents, getting rejected and re-uploading again. Let me explain!!

If you have ever taken policy or subscribed to one or planning to take one or you want to start some new account and all, you have to submit few documents. Luckily, scanned copies of documents uploaded on websites has minimized our visit to institutes/banks for above mentioned services. but there is one catch here.

Catch? Yes. Before explaining that let’s appreciate the fact that these online portals have saved lot of time for us. Catch is when we upload docs, those docs have be verified by someone who will see them on his/her monitor. so? If that person’s duty is over or he/she is not present today, verification of your docs is pending and you cannot move forward. Plus if docs get rejected and that person left for the day, even though you might upload correct doc in next hour but you are back to square one 😉

So ? what’s the solution. we need a way to verify our docs with less human interaction.

Let’s take an example of Aadhar card(which is like single card for identification) .Aadhar card has a number and if you get that number and query that number in database of govt records, you will get complete info of the person. so person who is on other side of monitor screen is doing that job(careful reader might think sumit why can’t we just manually enter number and let it verify? The thing is this eg is simple but there can be much more complicated workflow. This eg is taken to simplify end goal). so can we write a program that can help us detect that number in aadhar card and if we get that number, firing query, getting info and verifying user won’t be that big task.

How to detect number from that scanned image/pdf ? There is technique called OCR( Optical character recognition) and we can use deep learning to establish that technique.

I am going to show you a program that uses pretrained deep learning model and tries to predict. It is not perfect but with few more efforts, model can replicate human efforts in document verification.

I created fake identification card image(right side) and tried to extract text out if it. you can see model was able to extract text(left side in yellow ring) but order is little unusual.

so now we saw model is able to extract info. Based on this info, new info can be obtained and verification of user can be done efficiently with less human interaction.

For nerds, that pretrained model is combination of recurrent neural networks and convolution neural networks.

Hope you enjoyed reading this article and enjoyed your weekend as well(unlike me who was busy uploading docs 😓)