Search

Search Results

• ALGORITHM FOR TRAINING DATA PREPARATION OF CONVOLUTIONAL NEURAL NETWORKS FOR LETTER AND CHARACTER RECOGNITION

  D.А. Bezuglov , М.S. Mishchenko , S.E. Mishchenko

  134-144

  2025-07-24

  Abstract ▼

  The accuracy of text image recognition remains limited in practice. This is due to the fact that the alphabet of symbols can include lowercase and uppercase letters with a similar font, as well as composite characters formed from several simpler characters. To solve this problem, the character recognition system is supplemented with semantic or structural analysis systems, which significantly complicates the information system for text recognition. Currently, convolutional neural networks are widely used for recognizing single characters, for which a database with images of recognized characters is used for training. The paper proposes an algorithm characterized in that the image of a single character for a training sample includes fragments of characters that can be located in a line in close proximity to the recognized character.  This allows you to expand the set of images for training and additionally include information in the image about the placement of the symbol in the string, its relative size and whether this symbol is composite. The formation of images for the training sample simulates the process of segmentation of a symbol by brightness, which is usually used when selecting a symbol for further recognition.
  At the same time, the size of the symbol is estimated, it is supplemented with images of neighboring symbols, and then the size of the area, the image that will be placed in the training sample, is estimated. The resulting image is scaled and cropped in such a way that images of a given size are received at the input of the neural network. In the work, to recognize the alphabet of symbols, including uppercase and lowercase characters of the Russian and English alphabets, numbers, symbols and punctuation marks, it is proposed to use a variety of convolutional neural networks, each of which is trained to recognize one character. The symbol is selected by comparing the responses of all neural networks and selecting the maximum response. The proposed algorithm for training data preparation is compared with a well-known algorithm based on the use of images of single characters. It is established that the proposed algorithm for preparing data for training provides an increase in the accuracy of recognizing the alphabet of 138 characters by more than two times.