Natural scene text localization using edge color signature

Localizing text regions in images taken from natural scenes is one of the challenging problems due to variations in font, size, color and orientation of text. In this paper, we introduce a new concept so called Edge Color Signature for localizing text regions in an image. This method is able to localize both Farsi and English texts. In the proposed method first a pyramid using different scales of the input image is created. Then for each level of the pyramid an edge map is extracted. Afterward, several geometric features are employed to filter out the non-text edges from the extracted edges. At this stage we describe an edge using colors of its neighboring pixels. We use the mean-Shift algorithm to obtain the color modes surrounding each edge pixel. Subsequently, the connected edge pixels with similar color signatures are clustered using Single-Linkage clustering algorithm to construct meaningful groups. Finally, each of the clusters is labeled as text or non-text using an MLP based cascade classifier. The proposed method has been evaluated on well-known ICDAR 2013 and our Farsi dataset, the result is very promising.