Text Localization and Recognition in Real-Time Scene Images

Ananta Singh, Dishant Khosla

Abstract: An end-to-end text localization and detection is a much truer form of text extraction in image analysis if it can be harnessed efficiently. In this paper, a robust system to detect text from scene images is proposed. The text is extracted when image goes through multiple pre-processing stages. The method uses connected component (CC), Extremal Regions(ER) and sliding window concept. A set of rules and mathematical operations have been devised to extract actual text regions from background. Various techniques that are used to accurately localize text from scene images are implemented using HOG and stroke detection. Finally, text is recognized in an OCR stage of character recognition based on gradient direction feature and trained using synthetic fonts.

Keywords: localization, stroke detection, recognition, HOG transformation