1. Problem

1
1. Problem

• Many archived two-sided manuscript documents
  suffer from bleed-through
• Bleed-through can be effectively removed offline
  using image-processing algorithms
• A remotely located researcher may want to access
  both original and corrected versions of a
  document
• We want to avoid sending the document twice,
  since both versions are very similar.

Recto
Verso
2
3. Algorithm Details
Registration

• We assume that the continuous recto and verso
  image coordinate frames are related by a
  six-parameter affine transformation
• We search for a parameter vector that gives the
  best match between the recto and the transformed
  flipped verso, in the least-squares sense
• We identify the registered verso image

3
4. Joint Compression

• Based on existing standards
• Original, uncorrected image compressed with
  standard efficient compression scheme such as
  JPEG or JPEG 2000
• Segmentation map compressed using efficient
  bilevel compression scheme, such as JBIG or JBIG2
• Additional information for inpainting transmitted
  as side information

?
4.6Mbit
131 kbit
4
2. Bleed-through Removal
Model

• We assume the existence of underlying recto and
  verso images without bleed-though. These consist
  of the background,
  with the writing,
  superimposed.
• These ideal recto and verso images are combined
  in some way to produce the observed recto and
  verso images corrupted with bleed-through (see
  above).
• In general, the scanned recto and verso images
  (with bleed-through) will not be aligned.

Recto and flipped verso images superimposed
5
Segmentation

• We segment each side of the document into the
  four regions R1-R4. However, it is most important
  to correctly identify region R2, bleed-through
  only. If we miss some parts of R2, bleed-through
  will remain. If the label R2 is incorrectly
  assigned to some parts of R1, foreground only
  or R4 foreground and bleed-through, then parts
  of the desired writing will be erased.
• We first identify points that can be considered
  to definitely be background (R3), because they
  are lighter than a certain threshold.
• We then identify points that can be considered to
  foreground (R1), because they are darker than
  corresponding points on the other side.
• Of the remaining points, those whose correlation
  between the two sides exceeds a correlation
  threshold are deemed to be bleedthrough (R2). The
  rest are assigned to R4.

6
Original with bleed-through
With bleed-through removal
7
Algorithm

• Registration Alignment of recto and flipped
  verso
• Segmentation Four regions
• R1 Foreground only
• R2 Bleed-through only
• R3 Background
• R4 Foreground and bleedthroughoverlap
• Inpainting Region R2 filled in with estimate of
  background

Recto and flipped verso images, superimposed
after registration
Illustration of four types of regions
Inpainting applied to circled region
8
Inpainting

• Points labelled R2 bleed-through are replaced
  by suitable nearby points from the background
  region R3. In the initial work, a fixed value was
  used.

9
5. Conclusion

• Bleed-through can be effectively removed by
  jointly processing recto and verso sides of
  document.
• More complex bleed-through removal algorithms can
  be used at the server side, with the result
  transmitted to the remote user.
• It is not necessary to separately transmit
  original and corrected versions to a user who
  wishes to see both.
• All elements can be incorporated into JPEG2000.
• More work needs to be done on the segmentation
  and inpainting aspects of the algorithm.