Enhancing text image binarization using 3D tensor voting

Text image binarization is an important step in text image analysis and text understanding systems. Some corrupted regions may remain in the binarization result due to noises such as dust, streaks, shadows and small unwanted objects. In this paper, a novel method based on 3D tensor voting is proposed for enhancing text image binarization. The 3D tensor voting is used to detect corrupted regions by analysing surfaces of text stroke and background in a binary image. Our method is effective on binary images having gaps in text stroke or noise regions in background.