Scene text detection based on structural features

Abstract

While Optical Character Recognition (OCR) can be considered as a solved problem, text detection and recognition in real scene images is still extremely challenging and remains an open problem. Due to the wide variety of text appearances in real scenes, such as variations in font, size, color, orientation, partial occlusions, different distortions and illumination conditions, current results of both detection and recognition are still not satisfactory, as suggested by the low detection rates and recognition rates of state-of-the art approaches. One of the major reasons that degrade text detection and recognition accuracy is the large number of false positive characters, which is hard to handle due to diversity of text appearances in complex background. On the other hand, most of existing approaches do not provide an effective scheme to treat these false positive characters. In this paper, we propose a solution that takes into account the structural features of text strings. We prove these structural properties of text string can help obtain both removing false positive characters effectively and forming text-lines precisely. Experiments on the ICDAR Robust Reading Competition Dataset show a very competitive performance of our proposed approach.