Performance comparison of two text marking methods

摘要

A text document typically consists of a collection of regularnstructures such as words, lines, and paragraphs, a slight movement ofnwhich seems less perceptible than, say, dithering of the document image.nWe exploit this property to watermark formatted text documents bynshifting slightly certain lines and words in order to discourage illicitndistribution. We analyze two methods for reliable documentnidentification in the presence of severe distortions introduced bynphotocopying, facsimile transmission, and other processing. Thencorrelation method uses document profiles directly for detection. Toneliminate the effect of certain distortions, the centroid method basesnits decision on the distances between the centroids of adjacent profilenblocks. We present the maximum likelihood detectors for both methods andnevaluate their relative performance. Our analysis indicates thatnline-shift generally has a smaller error than word-shift detection, andnthat the correlation detector outperforms the centroid detector providedncertain distortions can be accurately compensated for before detectionnis attempted. These results have been applied to implement a marking andnidentification system and preliminary experimental results have beennvery promising