Neural network method for character sequence generation for text images training dataset synthesis

The size of the training sample is an important factor in solving optical character recognition tasks. Notably, the majority of the research focuses on increasing the variety of distributions that are applied to the images. Yet the internal structure of the textual information also affects the accuracy of the resulting model. We propose a neural network based text generation method for the creation of a synthetic training dataset of images with annotations, in which we propose to operate with groups of characters – alphabetic clusters, and use the sequence of clusters to predict the following character. The proposed cluster approach allows us to create specific sequences that retain the main properties of the target language, but do not contain a full language model. Since the proposed method works on a small number of clusters, we can use a small training set and a light neural network to generate text. The results of experiments with three open datasets of identity document images demonstrate the effectiveness of our method and the possibility of improving modern results for target fields.