Context: need to build an alternative model of recognition, to spend fewer resources on manual review of the documents.
Decision: a model for recognizing key document fields is built.
Results:
The solution is integrated into the customer's business process:
- Reduced staff costs 10 %;
- Improving the quality of risk analysis by 15 %;
- Increase the speed of application approval by 30 %.
Also, pilot tests were conducted on the following documents:
- Passport;
- SNILS;
- Driver's license;
- Passport of the vehicle.
To build the model, we used:
- Set of images and scans of personal documents;
- Marking up images with text boxes;
- The true value of each field;
- Database of named entities in the Russian Federation.
Simulation result:
- Model for recognizing the main elements of the passport;
- Model of character recognition for each element of the passport.
Customer: Finance, Banking, Insurance
Technology stack: OCR, TensorFlow, Python.