Problem Statement: Extracting information from PanCard is a continuous problem for the FinTech Companies, most of the companies these days are using Human Resources for this, but in this post Covid era all companies are looking for cots cutting and hence solution which can reduce human power and save cost is the need of the hour.
Lot of people tried this problem and prepared some heuristic solution, using OCR technology which doesn't promise accuracy more than 60-70%
Here we are trying to attempt a new Neural-Network based approach.
Proposed Methodology: With the help of advancement in object detection algorithms, we trained a neural-net model to detect area of useful information(bounding box), as shown in the image below.
This bound box will now help our OCR engine to search on the precise area with full energy, hence producing accuracy of more than 95%
Accuracy: Based on 300 testing images we observed the following accuracy-
Name: 94%
Father Name: 91%
DOB: 97%
PAN No: 96%
Model is taking just 1.5 Sec to predict the output, which is the minimum as compared to Microsoft API, Google Vision API and Karza API
Commercials of Existing Products:
Karza API : INR 4 per pancard, 74% accuracy
Microsoft : INR 3 per pancard, 78% accuracy
Google Vision:INR 3 per pancard, 79% accuracy
Here we developed complete package using open source
For further details on this use case please write to us at nakulrtm@gmail.com or call us at 6301415891
