Robust Local Thresholding Method for Segmentation of Non-Uniform Coloured Documents

Abstract

Text document segmentation is one of the essential steps in text document recognition and extraction systems. The existing image segmentation methods are not much reliable for text with colour gradient and texture background. Also, long processing time of existing methods is unfavorable. Thus, aiming to have reliable segmentation and less processing time, a new local thresholding method is proposed and its performance is tested in this study. The proposed method is based on pixel intensity magnification concept. In proposed algorithm, the input image is enhanced by edge sharpening. Then, the image is divided into multiple local windows by using non-overlapping. The magnification factor for each non-overlapping local window is calculated based on minimum intensity, maximum intensity, range and the number of dominant intensities in the corresponding local window. For segmenting pixels in each local window, magnification factor is used. The performance of proposed algorithm is measured in terms of segmenting accuracy and processing time. The tested images include different types such as text in uniform colour, text in multicolour, text in gradient background, text in national ID, text with watermark, highlight text, text in light illumination, text in passport, text in bank card. Also, the performances of Otus’ and Niblack’s methods are tested with the same images. The proposed method gives the better results for most images with maximum accuracy of 100% and lowest accuracy of 80%. The highest efficiency of Otsu’s method is 100% and its lowest accuracy is 0%. Since Otsu’s method loses all data sometimes. For Niblack’s method, it gives 100% accuracy only for text with simple colour background while its lowest accuracy is 10%. The proposed system’s average accuracy is higher than Otus’ and Niblack’s average accuracy.