Effective extraction of textual data from document images using transformer architecture of deep neural networks

V.A. Vykhodtseva, G.V. Popova, Y.A. Vais

Computer Optics, 2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.18287/coj1744
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

In the context of modern digital document management, the automation of document processing, particularly in accounting, is a crucial factor in enhancing the efficiency of business processes. However, automated document processing encounters a range of specific challenges, both linguistic and structural characteristics of the data. Traditional text processing methods that rely on classical optical character recognition (OCR) algorithms do not provide sufficient accuracy in extracting data from document images, which limits their use in automated accounting systems. These challenges are particularly evident when processing documents with complex structures, specific element placement, and text content. This paper proposes a solution to this problem by applying a model based on a transformer neural network architecture, specifically adapted for working with document images. Within the scope of this study, the transformer model is trained on a dataset of accounting document images with varying element placements and text with Cyrillic characters. The focus on Cyrillic text is particularly relevant, as research in this area has predominantly concentrated on documents in English or other Latin-based scripts. This article includes the results of training evaluated through specialized performance metrics. As a result of the experiment, at the final stage of training the model, the confidence loss was 0.156, which indicates that the model effectively minimizes the prediction error. The obtained accuracy of 0.868 showed a relatively high accuracy of forecasts. The Recall value of 0.905 indicates that the model effectively identifies most of the positive examples. The indicator F1=0.886 reflects a good balance between accuracy and memorability. The accuracy of 0.96798 indicates that the model's predictions are highly accurate. The use of the transformer model significantly improves the accuracy of extracting key information, such as date, number, and organization name, from accounting documents containing Cyrillic text. The findings of this study affirm the potential of this method for implementation in automated accounting systems, contributing to enhanced efficiency and precision in processing accounting documents.