Generative Adversarial Networks at Development of Automatic Document Processing System

The work is aimed to solving the applied problem of developing an automatic system for processing scanned / photo documents, based on machine learning approaches. The difficulty in developing systems of this type is the lack of unique examples to train the selected model. Therefore, to solve this problem, it is proposed to use generative adversarial neural networks. At this work was studied possibility of applying approach for artificial expansion of the training sample for further training of the selected model. The paper presents the classical architecture of the generator and discriminator, gives the number of parameters of the selected models and shows the time required to train these networks. Was making a study of the effects of the main parameters of generative adversarial networks on the quality of new examples generated by the generator. The results of this work show the possibility of applying the proposed approach in the development of systems of this type, for artificial, software expansion of the training sample by new, unique representatives.