DEVELOPMENT OF THE ARCHITECTURE OF DOCUMENT OPTICAL CHARACTER RECOGNITION SYSTEM

Abstract

This paper aims to develop information technology for document optical character recognition systems. The difficulty of processing images, which are a set of pixels, causes inconvenience in working with such data. This problem can be solved in different ways: usual mathematical approaches, a single complicated neural network, and a set of problem-specific deep neural networks. Usual mathematical approaches perform poor with unstructured data like images. A single neural network is like a swiss knife: it can do many tasks, but none with the best quality. So we will use two different deep neural networks, each for the appropriate part of the problem. The critical elements of this technology are the module for text detection and segmentation of the image, the module for text recognition in Ukrainian and English languages, the module for parsing multiple keywords, and the module for searching for the final data. The first and second modules consist of several machine learning models with specific architecture, depending on their task. All trained models are tested for accuracy and noise resistance and will be used in the future for searching required data from different document images. Output data of the developed system provide speedup, automation processing images and scans of the documents, reduce the number of mistakes caused by human factor. All data is converted from image pixels into a structured text set represented in the document, which the machine can easily use. We can use such technology in banking and insurance, where we can send images of documents and they will be automatically processed and converted into user name, surname, date of birth, serial number, and required fields for specific services.