Application of deep learning methods in the problems of text image segmentation

The paper is devoted to solving the problem of text image segmentation, the purpose of which is to select text blocks in the document image that correspond to columns, headers, footers etc. A review of existing image segmentation methods is carried out, including those intended for searching and selecting text blocks in images. Both classical methods and methods based on the use of artificial neural networks are analyzed. To solve given problem, an approach based on convolutional neural networks and the U-Net model is proposed. A method for automatically generating training examples for training a neural network is described. The processes of setting up a model, training and testing it are considered. The results of a numerical study of trained models on real data are presented.