A method for document image binarization based on histogram matching and repeated contrast enhancement

Paper in proceeding, 2014

In this paper, a new method for binarization of document images is introduced. During training, the method stores histograms from training images (divided into small tiles), along with the optimal binarization threshold. Training image tiles are presented in pairs, one noisy version and one clean binarized version, where the latter is used for finding the optimal binarization threshold. During use, the method considers the tiles of an image one by one. It matches the stored histograms to the histogram for the tile that is to be binarized. If a sufficiently close match is found, the tile is binarized using the corresponding threshold associated with the stored histogram. If no match is found, the contrast of the tile is slightly enhanced, and a new attempt is made. This sequence is repeated until either a match is found, or a (rare) timeout is reached. The method has been applied to a set of test images, and has been shown to outperform several comparable methods.