Neural networks and machine learning technologies are today at the peak of demand and popularity. They are used for text recognition, combined under common name – Optical Character Recognition (OCR). How they work, what documents can they recognize and why is this process worth trusting to neural networks?
Recognizing everything (but not immediately)
A neural network can recognize almost any text. But best of all the technology works with high-resolution printed texts – from 300 dpi. In this case, the document must be photographed or scanned exactly, there should be no illumination, disturbances, omissions and other distortions.
On recognition projects, our team often faces customers’ requests to digitize handwritten texts. They still meet in banking sector (various questionnaires, and so on), insurance sector, education (USE forms). Neural networks are not so easy to cope with them yet.
Technology has been improving over the past 5-6 years, but still neural networks are not so well-trained to recognize handwritten text with 100% accuracy. In addition, the text can be different, which also adds complexity.
For example, the system recognizes quite easily the handwriting of a design engineer who fills drawings manually, since it is close to printed text, contains the same letter spacing, and all letters are similar to the same font. But sometimes neural networks fail to recognize the handwriting of an ordinary person, since it contains a large number of connections between letters. As well as the famous “unreadable” doctors’ handwriting. Therefore, if a person cannot recognize visually what is written, then the neural network will not cope with this task either.
Once we decided to test one of the solutions for text recognition and scanned the page I wrote, and I have not the most legible handwriting. The result was – the system “broke,” the mission was impossible for it.
Why trust recognition to neural networks
They do it faster, make no mistakes and most importantly – reduce work required to process documents. For example, the accountant introducing manually data from the primary accounting documentation does it several times slower in comparison with solutions based on AI, and with a risk to make mistakes.
Neural network based recognition platforms can process up to hundreds of thousands of documents per day. Volumes depend on capacity of computing resources.
Not only native integrations, but also software robots are used for maximum process acceleration and data migration. They help transfer data automatically from the recognition system to the company’s accounting system or electronic archive.
In addition, robots can check documents for authenticity. From the beginning, neural network recognizes the counterparty’s card in the contract, and then the robot checks data with different resources or other documents.
What neural networks can do
In addition to simple recognition for further data transfer to other systems, neural networks are capable of:
- recognize the presence of artifacts (stamps, seals) for understanding legal strength of the document;
- search for entities and check their relevance;
- determine the completeness of the document by attributes. The system helps to check the legal force of the entire set and its integrity determines the type of documents, searches for instances that require manual processing and performs cross-processing verification.
In addition, neural network based solutions can extract metadata and entities from the document and check them against reference books.
How does it work
The first step is to enter documents. This could be uploading electronic versions from “hot” folder to the system or paper copies to the scanner. On this step you can check the completeness of the kit. If some document is missing, the system itself will ask to load it.
The second step is document processing. Neural network extracts data, checks availability of necessary attributes, correctness of calculations, etc. Step three is data verification. At this stage, the correctness of all information is checked.
If the neural network does not fully recognize the document, it sends data to a person for manual adjustment.
The fourth step is to export documents. For example, to the accounting system or archive.
Cases from practice
The customer generated about 100 thousand documents per month in the general service center. For processing, a system for recognizing whole document flow was used, the process was executed directly on the server, not controlled by users and there was no verification step. For transferring documents to the archive, it was necessary to check how their cards were filled out. The key users of the system were warehouse employees, accounting, HR, administration, and individual management groups.
After the introduction of the new solution, the process was built differently. Now 95% documents fall into the archive without recognition, since first the system reads the barcode and QR code of document packages. Next recognition occurs according to pre-prepared templates. Those documents that require the attention of a specialist are sent for verification .
The processing time of documents was reduced from five minutes to one minute per one document. If earlier it was necessary to open both a card and documents and check with unrecognized ones, now it is done automatically. The costs were also optimized.
In one of the universities, it was necessary to introduce a recognition system for analysis of the presence of signatures and seals in the documents. Previously educational institution used a solution that simply registered documentation and uploaded to the archive without verification. Because of this, some of the documents were archived without the necessary attributes, which further led to controversial situations with counterparties and long searches for the necessary form.
During the project, a neural network was created to analyze the presence of signatures and seals in documents. Also, the entire previously formed archive passed the check. The recognition accuracy is now about 96%.
In addition, software robots are used to work with documents.
The labor costs of university employees decreased by 5 times, and the risk of the occurrence of any disputes with counterparties is minimized.
Despite the fact that now all companies are trying to move away from paper document management, recognition systems are still relevant and needed.
Neural networks help businesses speed up workflows, process and check documents. And since the systems are constantly improving, then in in the future, we are waiting for an increase in the percentage of constant recognition accuracy, including handwritten texts.
By Vladislav Chernetsky, IT Business Development Manager, Konica Minolta Business Solutions Russia