Motivation for the customer to launch the project:
the need to scale the customer’s business revealed limitations in the current solution: the business model of the product used did not allow reducing the cost of the solution when scaling.
Description of the initial situation:
Project goals:
MIL Team solution:
the use of the team’s existing solutions in the field of detection and recognition of text in images based on neural network models made it possible to quickly implement a solution with the required level of quality.
To build the model we used:
Simulation results:
Full name recognition accuracy >85%
Gender recognition accuracy >88%
Recognition accuracy of license plates and series >91%
Accuracy of place and date of birth recognition >73%
Accuracy of recognition of the Authority and date of issue >80%
Customer: Finance, Banking, Insurance
Technology stack: OCR, TensorFlow, Python.
the need to scale the customer’s business revealed limitations in the current solution: the business model of the product used did not allow reducing the cost of the solution when scaling.
Description of the initial situation:
- checking a client’s creditworthiness when issuing a bank loan involves;
- recognizing his personal identification documents using semi-automatic and automatic methods;
- reducing the risk of incorrect document recognition is ensured by high-quality cross-checking of documents by operators;
- the manual and automated solutions used had scaling limitations for this reason, there were significant risks for the planned business growth.
Project goals:
- creation of a recognition model for the main fields of identification documents with a sufficient level of quality;
- creating a model for a fixed cost, which will avoid an inappropriate business model (transactional) when scaling.
MIL Team solution:
the use of the team’s existing solutions in the field of detection and recognition of text in images based on neural network models made it possible to quickly implement a solution with the required level of quality.
To build the model we used:
- A set of images and scans of personal documents;
- Marking images with text boxes;
- The true value of each field;
- Database of named entities of the Russian Federation.
Simulation results:
- Model for recognizing the main elements of a passport;
- Character recognition models for each passport element.
Full name recognition accuracy >85%
Gender recognition accuracy >88%
Recognition accuracy of license plates and series >91%
Accuracy of place and date of birth recognition >73%
Accuracy of recognition of the Authority and date of issue >80%
Customer: Finance, Banking, Insurance
Technology stack: OCR, TensorFlow, Python.
