Seite - 12 - in Document Image Processing

J. Imaging 2018,4, 68 Figure3.Visualcomparisonofbleed-throughrestoration. (a) inputdegradedimage; (b)handlabeled ground truth image; (c) restored imageby [16]; (d) restored imageby [7]; (e) restored imageby the proposedmethod. Figure4. Inkbleed-throughremoval inacolour image: input image(top)andtherestoredimageusing theproposedmethod(bottom). 6.Conclusions This paper presents a novel and general framework for high quality image restoration of documents affectedbybleed-through. Weuse thebleed-through identificationmethodpresented in [33] in conjunctionwith group based sparse image inpainting, in order to obtain a non-blind document bleed-through the removalmethod. Thenon-stationary linearmodel in [33] efficiently locates thebleed-throughpattern in recto-verso imagepairs, but lacksapropermethod to replace theunwantedbleed-throughpixels. Finding abefittingfill-in for thedegradedpixels is a crucial 12