Contrast-Aware Halftoning
Hua Li and David Mould
April 20, 2023 1
Previous Work
April 20, 2023 2
Original Image Floyd-Steinberg error diffusion[FS74]
Tone Reproduction
Visual artifacts
Lack of structure preservation
Previous Work
April 20, 2023 3
Ostromoukhov’s method[Ost01]
Tone ReproductionBlue Noise
Floyd-Steinberg error diffusion[FS74]
Visual artifacts
Lack of structure preservationLack of structure preservation
Improved
Previous Work
April 20, 2023 4
Ostromoukhov’s method[Ost01] Structure-aware halftoning[Pang et al. 2008]
Blue Noise Structure Preservation
Structure preservation
Very slow
Lack of structure preservation
Previous Work--Current Art of State
April 20, 2023 5
Structure-aware error diffusion[Chang et al. 2009]
Structure PreservationStructure Preservation
Structure preservation
Very fast but a little lower quality in structure preservation
Structure preservation
Very slow
Structure-aware halftoning[Pang et al. 2008]
Comparison with Our Work
April 20, 2023 6
Contrast-aware halftoning(Our variant method) Structure-aware halftoning[Pang et al. 2008]
Motivation• Human perception is sensitive to contrast.• Visual effect/impression more important than
tone matching.• Observation(at the core of our algorithm)– Using more black pixels in the dark side and fewer
black pixels on the light side will promote the local contrast.
April 20, 2023 7
Observations for Contrast Enhancement
April 20, 2023 8Artists’ work
Goal and Problem
• Goal: Structure preservation without loss of tone quality and sacrificing speed
• Problem:– How to cluster black pixels in white area to
maintain local contrast for generating structure-preserved monochrome halftoning ?
April 20, 2023 9
1. Our Basic Algorithm
• Basically, our basic method is an extension to Floyd-Steinberg error diffusion.– Pixel by pixel
April 20, 2023 10
Contrast-aware mask
p(i,j)
1. Our Basic Algorithm
April 20, 2023 11
1. Determine the pixel color: (closer to black) or (closer to white);
2. Calculate the error(the difference): the original intensity - the chosen intensity;
3. Calculate the weights of contrast-sensitive mask;4. Normalize the weights;5. Diffuse the error.
For each pixel p(i,j)
Based on FS error diffusion
Contrast-preserved Error Distribution
April 20, 2023 12
255
0
0
128
255
128
<128
0
The center pixel The center pixel
Positive error
>128Negative error
255
p(i,j)
Nearby pixels Lightened
Nearby pixels Darkened
Uniform Region
Contrast-preserved Error Distribution
April 20, 2023 13
255
255
0
0
Positive error
Negative error
Original After
Non-uniform Region
Contrast-preserved Error Distribution
• Contrast-sensitive circular mask– Maintain the initial tendency that darker pixels should
be more likely to be set to black while lighter pixels should be more likely to be set to white. • The nearby darker pixels absorb less positive error and the
lighter pixels absorb more.• Conversely, negative error is distributed preferentially to
dark pixels, making them even darker.
– Weights steeply dropping off from center– Normalized
April 20, 2023 14
Comparisons for Ramp
April 20, 2023 15
Ostromoukhov’s method
Structure-aware halftoning
Our basic method(Have annoying patterns)
Floyd-Steinberg error diffusion
Ramp
2. Our Variant Method
• Instead of the raster scanning order, dynamically priority-based scheme– Closer to either extreme(black or white), higher
priority.
April 20, 2023 16
Contrast-preserved Error Distribution
April 20, 2023 17
255
0
0
128
255
128
<128
0
The center pixel The center pixel
Positive error
>128Negative error
255
p(i,j)
Uniform Region
Highest priority
Highest priority
Highest priority
Highest priority
Lowered
Lowered
Priority-based Scheme
• The neighboring pixels change priorities after using contrast aware mask.
• The neighboring pixels will not be chosen as the next pixel. To guarantee a better spatial distribution.
• An up-to-date local priority order, empirically, results in superior detail preservation.
April 20, 2023 18
Visualize the Orders after Our Variant method
April 20, 2023 19
Visualize the orders for the tree image. - The first pixel is set as black and the last pixel is set as white.
Comparisons for Ramp
April 20, 2023 20
Our basic method(Have annoying patterns)
Our variant method
Improvement for Mid-tone
April 20, 2023 21
Ostromoukhov’s method
Structure-aware halftoning
Our variant method
Floyd-Steinberg error diffusion
Ramp intensity
Part of Tree
April 20, 2023 22
(a)Structure-aware halftoning (b)Structure-aware error diffusion
(c)Our basic method (d)Our variant method
Snail
April 20, 2023 23
April 20, 202324
Structure-aware halftoning
Structure-aware error diffusion
Our basic method
Our variant method
Comparisons(1)
April 20, 2023 25
April 20, 2023 26
SAH
SAED
Basic
Variant
Comparisons(2)
April 20, 2023 27
April 20, 2023 28
Comparisons(4)
April 20, 2023 29Structure-aware halftoning Our basic method Our variant method
Evaluation for Structure Similarity
April 20, 2023 30MSSIM(the mean structural similarity measure[Wang et al. 2004])
EvaluationTone Similarity and Structure Similarity
April 20, 202331
The peak signal-to-noise ratio(PSNR)
MSSIM
Evaluation-Contrast Similarity
April 20, 2023 32
the peak signal-to-noise ratio based on local contrast image(CPSNR)
Blue Noise Properties by the Radially Averaged Power Spectrum
April 20, 2023 33
Grayness = 0.82Our basic method and its RAPSD Our variant method and its RAPSD
Our variant method with tie-breaking and its RAPSD
Structure-aware method and its RAPSD
Analysis
• CPU Timing(Process a 512 ×512 image)
• Limitation: not optimal; sometimes clumping happens.
April 20, 2023 34
Methods Structure-aware halftoning
Structure-aware error diffusion(16×16 mask)*
Our basic method(7×7 mask)**
Our variant method(7×7 mask)**
Time 2 minutes 6.74 seconds 0.492 seconds 2.955 seconds
* Best tradeoff between quality and speed** Similar hardware conditions as SAED
Summary
• We have a tradeoff of intensity fidelityvs. structural fidelity and have the best structure preservation of any reported results to date.
• Contrast-aware halftoning is automatic, easy to implement, and fast.
• Contrast is an important factor.
April 20, 2023 35
Contributions
• Based on error diffusion, propose contrast-aware methods for halftoning creation.
• Introduce dynamically priority-based scheme into halftoning.
April 20, 2023 36
Future Work
• Shape influences• Other image features to adjust local contrast• Color halftoning• Other artistic styles through pixel
management
April 20, 2023 37
Acknowledgement
• Thanks to:
Grants from NSERC and Carleton University
April 20, 2023 38
More Results:Based on Our Variant Method
April 20, 2023 39
April 20, 2023 40
April 20, 2023 41
April 20, 2023 42
April 20, 2023 43
April 20, 2023 44
April 20, 2023 45
April 20, 2023 46
April 20, 2023 47
April 20, 2023 48
April 20, 2023 49
April 20, 2023 50