Text Localization in Unconstrained Images

Abstract. Text localization is the first step when au-tomatically reading text in images. Since existing methods often fail when applied to unconstrained im-ages, in this paper we propose a more robust ap-proach exploiting different kind of information. In particular, we first extract textural features, a com-bination of a stroke filter with a super-pixel segmen-tation, and then search for connected components. To finally obtain a text localization, these are subse-quently analyzed for unary character properties, bi-nary character similarities, and text line properties. To demonstrate the benefits of the proposed method, we evaluate it on three different data sets, showing promising results. 1