Extension for Interactive Approach
Ink‐Bleed Reduction Using Functional Minimization
Grani A. Hanasusanto
Zheng Wu Michael S. Brown
• For some documents, the foreground and ink‐bleed intensities may vary spatially
School of Computing National University of Singapore
Introduction
• Ink‐bleed reduction algorithm applied globally will generate inaccurate foreground l b ll ill
i
f
d
segmentation over the local region
•Our proposed interactive approach is to let user denote the region where local minimization is to be performed
Our approach
Ink‐bleed problem
• Ink‐bleed removal by a modified Chan‐Vese active contour model, taking into account the information from both recto and verso image • Ink seeps through a paper document and interferes with the text written on the opposite side
Previous approaches
• Methods based on wavelet, MRF, blind signal separation, diffusion model etc, usually involving a set of parameters to be adjusted or estimated based on prior assumptions
set of parameters to be adjusted or estimated based on prior assumptions
• Functional minimization for completing broken strokes that arise when strong ink‐bleed overlaps the foreground strokes
p
g
• classification methods need user‐supplied markup as training data, either requiring sufficient markup or resorting to directed assistance for iterative marking up, but independent of parameters • Interactive tool to improve the results over local regions, for documents with non‐uniform characteristics
•The resulting minimization is optimal as it takes into account the statistics in the locality
Front image
Workflow
Inputs
Ink‐Bleed Reduction Stroke Completion +
Front image
Cleaned front image result
Local minimized result (in red)
Initial global result
Zoomed regions
Zoomed regions Global results
Final result
Back image
Local results
Completion domain in green
Ink Bleed Reduction Algorithm
Ink‐Bleed Reduction Algorithm
Energy
recto image u 0
background
recto image
u0
fg - fg / fg - fg
2
4
3
1
foreground
+ ink-bleed
c1
verso image
c2
-128
255
1
2
3
4
3/4
min(u0 − v0 ,0)
v0
0
Stroke Completion
Energy E 2
E1
fg - fg
fg - fg
1
2
c3
0
1
2
3
4
u0
u0
u0
u0
< v0
> v0
≈ v0
≈ v0
128
Experimental Results
i
l
l
• When ink‐bleed is severe, the parameter λ
needs to be increased considerably, resulting in the broken strokes
• Perform statistical significance testing to determine the domain in need of completion • Use modified Cahn‐Hilliard functional to Use modified Cahn Hilliard functional to
inpaint the broken stroke in the completion domain
recto foreground & verso non‐foreground pixel pair ( ) fg - fg
recto non‐foreground & verso foreground pixel pair ( ) fg - fg
recto non‐foreground & verso non‐foreground pixel pair ( ) fg - fg
recto foreground & verso foreground pixel pair ( ) fg - fg
Result of Individual and Combined Energies
• Minimize E1 to remove the background, but the ink‐
bleed cannot be discriminated from the foreground
u0
E1
C1
• Minimize E2 to extract the pixel pairs, but the fg - fg
both‐foreground pixel pairs are removed
• Minimize the combined energy functional
E = length
(C ) + E 1 + λ E 2
to generate the desired result, where λ is a user defined parameter for controlling the balance between energy E1 and E2
C
E1 + λ E 2
u 0 − v0
C2
E2
Inputs
Initial results
Completion
domains
Final results
Input
Classification based results
Our results