The authors of this paper combine global and local methods for detecting faint characters, bleed-through, and large background ink stains, and propose an adaptive document image binarization method applied at the connected component level. They classify document degradations as shadows, nonuniform illumination, and smudges that affect the text and are carried as noise in the images. The method estimates the document image background with image normalization, followed by simultaneous global and local binarization methods to compute stroke width and contrast, and then combines the resulting binarized images.
展开▼