A method for text localization and recognition in real-world images

A general algorithm for text detection and recognition in real-world images is presented in this paper, which due to numerous problems is a harder task than standard printed document recognition. The algorithm finds text areas in photographs taken by a standard camera or a mobile phone and `reads' content of the detected text areas, even when the text occupies just a small part of the image, when the camera is not aimed directly at the text and when lighting conditions are not perfect. The algorithm uses many innovative pieces, such as simultaneous processing of multiple text line hypothesis, feedback loops to correct wrong decisions of previous steps, use of synthetic fonts to train the algorithm (so that there is no need for time-consuming acquisition and labeling of real-world training data) or character recognition based on line context using a typographic model. The algorithm was tested on four datasets of real-world images, from which two datasets are public and they have been already used to evaluate performance of other existing methods for text detection and recognition. The proposed algorithm outperforms the other existing methods by reading correctly 71\% respectively 60\% characters in these two datasets, thus achieving the state-of-the-art results. We conclude by showing possible applications of the proposed algorithm, such as automatic indexing of images with text into a database, automatic information retrieval for mapping applications, mobile phone application to help blind people, automatic system for drivers that warns about surrounding traffic signs or an automatic translator of foreign signs and labels. The algorithm was also successfully tested on cyrillic text.

In Czech.