A method for binarization of document images from a live camera stream

Paper in proceeding, 2015

This paper describes a method for binarization of document images from a live camera stream. The method is based on histogram matching over partial images (referred to as tiles). A method developed previously has been applied successfully to images with artificially added noise. Here, an improved method is presented, in which the user has more direct control over the specification of the binarizer. The resulting system is then taken a step further, by considering the more difficult case of binarization of live camera images. It is demonstrated that the improved method works well for this case, even when the image stream is obtained using a (slightly modified) low-cost web camera with low resolution. For typical images obtained this way, a standard OCR reader is capable of reading the binarized images, detecting around 87.5% of all words without any error, and with mostly minor, correctable errors for the remaining words.