Scene text localization based on the ultimate opening.
Thomas Retornaz, Beatriz Marcotegui · Biblioteca Digital da Memória Científica do INPE (National Institute for Space Research) · 2007
Database indexing and retrieval tools can benet enormously from the automatic detection and processing of textual data in images. We present a new connected-component based method for automatic localization of text in general image database. This technique is robust to font, scale and slant changes and detects superimposed as well as scene text. We rst use a non linear scale space based on morphological numerical residue (ultimate opening). From this we extract candidate CC using contrast information. We then use a two-step lter to eliminate incorrect CCs: rst by using spatial coherence of text zone to eliminate noise CC then by cor- relation between global geometry and thickness of characters. We then classify the remaining CCs by such criteria as global geom- etry, thickness consistency, contrast and shape regularity. A nal aligning and merging analysis is performed in order to lter out the remaining non-text CCs and to compute the nal bounding boxes of the text zones. The robustness of our approach is proven by the results in the ImagEval evaluation campaign, which database included old postcards, graphic schemes, stamps, indoor and out- door scene images and also images without any textual data. In spite of the wide variety of texts and images our approach obtains interesting results without parameter tuning for each image class.