A framework for reducing ink-bleed in old documents

Yi Huang, Michael S. Brown, Dong Xu

Abstract: We describe a novel application framework to reduce the effects of ink-bleed in old documents. This task is treated as a classification problem where training-data is used to compute per-pixel likelihoods for use in a dual-layer Markov Random Field (MRF) that simultaneously labels image pixels of the front and back of a document as either foreground, background, or ink-bleed, while maintaining the integrity of foreground strokes. Our approach obtains better results than previous work without the need for assumptions about ink-bleed intensities or extensive parameter tuning. Our overall framework is detailed, including front and back image alignment, training-data collection, and the MRF formulation with associated likelihoods and intra- and interlayer cost computations.