A statistical tool based binarization method for document images

摘要

Binarization of document images has great importance in several applications like historical document restoration, Optical Character Recognition (OCR). It is a challenging task due to small difference between foreground and background pixel intensities, intricate font patterns and noisy background. In this article a binarization algorithm is presented for document images which has performed significantly well on handwritten document images as well as machine printed document images. First, the RGB document images are converted to a prominent gray-scale image using statistical tools like mean, variance and standard deviation. Next, the gray-scale images are binarized using edge detection. Further the noises are removed using connected component features analysis. The proposed method is experimented on publicly available DIBCO 2016 and DIBCO 2017 datasets. The performance of the proposed algorithm is satisfactory in terms of F-Measure (FM), Pseudo-FMeasure (F-ps), PSNR, Distance Reciprocal Distortion (DRD) and it also provides significant results on degraded document images.