Creating Training Datasets For OCR In Mobile Device Video Stream

This paper studies methods of data sampling for training of convolutional neural networks for character recognition. These methods are considered for optical character recognition of machine readable zone (MRZ) of documents captured by a mobile phone camera. Advantages and disadvantages of training on natural and artificial datasets are discussed. In this paper we describe some set of image transformations and give examples of their practical implementation. At the end we show how adding artificial examples (to the training database), generated according to the analysis of recognition error, improves the quality of recognition.