Color Stability and Homogeneity Regions to Detect Text in Real Scene Images: CSHR

Abstract

In this paper, a novel method called CSHR for affine invariant detection of stable and homogeneous parts of the extremal regions to localize text in natural scene images is proposed. The basic idea of this method was to apply two local thresholds to extract the extremal regions by their color homogeneity and select the candidate regions by maximum and minimum surface limits. Then, the candidate regions were filtered according to a stability criterion to extract the maximally stable parts of the extremal regions. Finally, the text regions are filtered using region area, orientation, and aspect ratio properties as well as features specific to the Arabic language to focus on Arabic writing. The proposed approach which was tested on the ICDAR2003 database and on our database showed an improvement over the existing methods.