An Adaptive Threshold Method for Gray Scale Character Images

Gray-scale document images are binarized in order to perform optical character recognition (OCR). To perform this binarization, a variety of techniques have been proposed for performing threshold selection. Most of these are based on the intensity distribution of the image regions. However, when the spacing between two lines is very small, it is difficult to produce high quality characters using a fixed threshold. Also, with a fixed threshold, it is very hard to produce continuous lines when bina-rized, because the binarization introduces gaps within the line. This paper presents an adaptive threshold method based on gradient properties which handles the above problems. Our method consists of three steps. First we extract the pixels located at the boundary of a character and the background, then determine a threshold for each of the extracted pixels. After binarizing the boundary pixels, the remaining pixels are binarized based on the binarized boundary pixels. Our results show that an adaptive threshold produces higher quality binarized images than does the discriminant analysis method.