Generation of Synthetic Images of Full-Text Documents

摘要

In this paper, we present an algorithm for generating images of full-text documents. Such images can be used to train and evaluate models of optical character recognition. The algorithm is modular, individual parts can be changed and tweaked to generate desired images. We describe a method for obtaining background images of paper from already digitalized documents. We use a Variational Autoencoder to train a generative model of these backgrounds enabling the generation of similar background images as the training ones on the fly. The module for printing the text uses large text corpora, font, and suitable positional and brightness noise to obtain believable results. We use Tesseract OCR to compare the real world and generated images and observe that the recognition rate is very similar indicating the proper appearance of the synthetic images. Furthermore, the mistakes made by the OCR system in both cases are alike. Finally, the system generates detailed, structured annotation of the synthesized image.