Comparison of some thresholding algorithms for text/background segmentation in difficult document images

Abstract

A number of techniques have previously been proposedfor effective thresholding of document images. In this papertwo new thresholding techniques are proposed andcompared against some existing algorithms.The algorithms were evaluated on four types ofdifficult' document images where considerablebackground noise or variation in contrast and illuminationexists. The quality of the thresholding was assessed usingthe Precision and Recall analysis of the resultant words inthe foreground.The conclusion is that no single algorithm works well forall types of image but some work better than others forparticular types of images suggesting that improvedperformance can be obtained by automatic selection orcombination of appropriate algorithm(s) for the type ofdocument image under investigation.