User-assisted ink-bleed reduction

Abstract

This paper presents a novel user-assisted approach to reduce ink-bleed interference found in old manuscripts. The problem is addressed by first having the user provide simple examples of foreground ink, ink-bleed, and the manuscript's background. From this small amount of user-labeled data, likelihoods of each pixel being foreground, ink-bleed, or background are computed and used as the data costs of a dual-layer Markov random field (MRF) that simultaneously labels all pixels in both the front and back sides of the manuscript. This user-assisted approach produces better results than existing algorithms without the need for extensive parameter tuning or prior assumptions about the inkbleed intensity characteristics. Our overall application framework is discussed along with details of the features used in the data costs, a comparison between -nearest neighbor and support vector machine for likelihood estimation, the dual-layer MRF formulation with associated inter- and intra-layer costs, and a comparison of our approach against other ink-bleed reduction algorithms. © 2010 IEEE.