An Efficient Detection Method for Text of Arbitrary Orientations in Natural Images

Abstract

Due to the high complexity of natural scenes, text detection is always a critical yet challenging task. On the basis of existing character detection method, a novel text line detection method is proposed in this paper, which can localize text of arbitrary orientation by using related information of character regions in candidate text line. First, inspired by the Hough transform, text line detection problem is regarded as line detection problem in candidate characters set obtained by Most Stable Extremal Regions (MSERs). Second, in order to find out the relationship of adjacent candidate regions, a graph model is built based on some constraints and adjacent candidates are linked into pairs to obtain search domain. Then, to avoid repeated calculation of the same line, some strategies need to be used. Finally, as some of the potential text lines are incorrect, we use a new text line descriptor to exclude the non-text areas. Experimental results on the ICDAR 2013 competition dataset and MSRA-TD500 show that the proposed approach is favorable no matter for non-horizontal text or horizontal text.