Segmentation of unstructured text on the book cover images using the convolutional network based on the U-Net architecture

Abstract

This paper discusses the convolutional neural network for image segmentation with book covers. The structure of the network is given, indicating all its constituent blocks and layers, as well as their parameters, and the operating principle of each part is described in detail. The U-Net model is used as the basis of the network. The architecture of this model stands out among others with its encoder-decoder structure, which allows generating new images. In this case, the encoder part of the network is responsible for image recognition, and the decoder part is responsible for generating a new image. The proposed neural network is capable of creating binary (black and white) masks, on which the text is highlighted in one color, and all other elements in another. Thus, the text is separated from other elements in the image. To train and test the convolutional neural network, the self-assembled and labeled dataset of 200 examples is used. Despite the small amount of data, the U-Net-based network trains well and shows acceptable performance results, which is confirmed by the test results. The trained network can be used in practice. In particular, it is supposed to be used to improve the accuracy of text recognition on book covers.