An improved approach for automatic denoising and binarization of degraded document images based on region localization

This paper presents a new binarization approach in which the binarization algorithm is neither applied on whole image as in case of global thresholding nor on sub images as in case of local thresholding. In this approach, using the concept of bounding box and edge detection method, region localization is done and the objects of interest i.e. the textual part is separated from the background part of the image. After the extraction of object of interest, local binarization is applied to only that region using the overlapping windows. The result of this approach is then compared with the result of other local binarization algorithms like midgrey, sauvola, niblack, mean etc. In this experiments were done on different datasets including DIBCO2009, DIBCO2010, DIBCO2011, DIBCO2012 and DIBCO2013. Basic evaluation parameters used for comparison are precision, recall, F-measure, mean square error (MSE), signal to noise ratio (SNR) and peak signal to noise ratio (PSNR).