A skeleton based binarization approach for video text recognition

Text in video data comes in different resolutions and with heterogeneous background resulting in difficult contrast ratios that most times prohibit valid OCR (Optical Character Recognition) results. Therefore, the text has to be separated from its background before applying standard OCR process. This pre-processing task can be achieved by a suitable image binarization procedure. In this paper, we propose a novel binarization method for video text images with complex background. The proposed method is based on a seed-region growing strategy. First, the text gradient direction is approximated by analyzing the content distribution of image skeleton maps. Then, the text seed-pixels are selected by calculating the average grayscale value of skeleton pixels. And finally, an automated seed region growing algorithm is applied to obtain the text pixels. The accuracy of the proposed approach is shown by evaluation.