C. A.

Bouman: Digital Image Processing - January 8, 2018 1

Digital Halftoning
• Many image rendering technologies only have binary out-
put. For example, printers can either “fire a dot” or not.
• Halftoning is a method for creating the illusion of contin-
uous tone output with a binary device.
• Effective digital halftoning can substantially improve the
quality of rendered images at minimal cost.
C. A. Bouman: Digital Image Processing - January 8, 2018 2

Thresholding
• Assume that the image falls in the range of 0 to 255.
• Apply a space varying threshold, T (i, j).

255 if X(i, j) > T (i, j)
b(i, j) = .
0 otherwise
• What is X(i, j)?
• Lightness
– Larger ⇒ lighter
– Used for display
• Absorptance
– Larger ⇒ darker
– Used for printing
• X(i, j) will generally be in units of absorptance.
C. A. Bouman: Digital Image Processing - January 8, 2018 3

Constant Threshold
• Assume that the image falls in the range of 0 to 255.
• 255 ⇒ Black and 0 ⇒ W hite
• The minimum squared error quantizer is a simple thresh-
old 
255 if X(i, j) > T
b(i, j) = .
0 otherwise
where T = 127.
• This produces a poor quality rendering of a continuous
tone image.
C. A. Bouman: Digital Image Processing - January 8, 2018 4

The Minimum Squared Error Solution

• Threshold each pixel
– Pixel> 127 Fire ink
– Pixel≤ 127 do nothing
Original Image Thresholded Image

50 50

100 100

150 150

200 200

250 250

300 300

350 350

50 100 150 200 250 50 100 150 200 250

C. A. Bouman: Digital Image Processing - January 8, 2018 5

Ordered Dither
• For a constant gray level patch, turn the pixel “on”in a
specified order.
• This creates the perception of continuous variations of
gray.
• An N × N index matrix specifies what order to use.
 
1 2
I2(i, j) =
3 0
• Pixels are turned on in the following order.

0 1 2 3 4
C. A. Bouman: Digital Image Processing - January 8, 2018 6

Implementation of Ordered Dither via

Thresholding
• The index matrix can be converted to a “threshold matrix”
or “screen” using the following operation.
I(i, j) + 0.5
T (i, j) = 255
N2
• The N × N matrix can then be “tiled” over the image
using periodic replication.
T (i modN, j modN )

• The ordered dither algorithm is then applied via thresh-

olding.

255 if X(i, j) > T (i modN, j modN )
b(i, j) = .
0 otherwise
C. A. Bouman: Digital Image Processing - January 8, 2018 7

Clustered Dot Screens

• Definition: If the consecutive thresholds are located in
spatial proximity, then this is called a “clustered dot screen.
• Example for 8 × 8 matrix:
62 57 48 36 37 49 58 63
56 47 35 21 22 38 50 59
46 34 20 10 11 23 39 51
33 19 9 3 0 4 12 24
32 18 8 2 1 5 13 25
45 31 17 7 6 14 26 40
55 44 30 16 15 27 41 52
61 54 43 29 28 42 53 60
C. A. Bouman: Digital Image Processing - January 8, 2018 8

Example: 8 × 8 Clustered Dot Screening

Cluster Dot Screen of Size 8

50

100

8x8 Cluster Dot 150

2
200

250
4

5
300
6

7 350

1 2 3 4 5 6 7 8 50 100 150 200 250

• Only supports 65 gray levels.

C. A. Bouman: Digital Image Processing - January 8, 2018 9

Example: 16 × 16 Clustered Dot Screening

Cluster Dot Screen of Size 16

50

100

16x16 Cluster Dot 150

200
4

6
250
8

10 300

12

14
350

16
2 4 6 8 10 12 14 16 50 100 150 200 250

• Support a full 257 gray levels, but has half the resolution.
C. A. Bouman: Digital Image Processing - January 8, 2018 10

Properties of Clustered Dot Screens

• Requires a trade-off between number of gray levels and
resolution.
• Relatively visible texture
• Relatively poor detail rendition

• Uniform texture across entire gray scale.

• Robust performance with non-ideal output devices
– Non-additive spot overlap
– Spot-to-spot variability
– Noise
C. A. Bouman: Digital Image Processing - January 8, 2018 11

Dispersed Dot Screens

• Bayer’s optimum index Matrix (1973) can be defined re-
cursively.
 
1 2
I2(i, j) =
3 0
 
4 ∗ In + 1 4 ∗ In + 2
I2n =
4 ∗ In + 3 4 ∗ In
• Examples
21 37 25 41 22 38 26 42
53 5 57 9 54 6 58 10
5 9 6 10 29 45 17 33 30 46 18 34
1 2 13 1 14 2 61 13 49 1 62 14 50 2
3 0 7 11 4 8 23 39 27 43 20 36 24 40
15 3 12 0 55 7 59 11 52 4 56 8
31 47 19 35 28 44 16 32
63 15 51 3 60 12 48 0

2×2 4×4 8×8

• Yields finer amplitude quantization over larger area.
• Retains good detail rendition within smaller area.
C. A. Bouman: Digital Image Processing - January 8, 2018 12

Example: 8 × 8 Bayer Dot Screening

Bayer Screen of Size 8

50

100

8x8 Bayer Dot 150

2
200

250
4

5
300
6

7 350

1 2 3 4 5 6 7 8 50 100 150 200 250

• Again, only 65 gray levels.

C. A. Bouman: Digital Image Processing - January 8, 2018 13

Example: 16 × 16 Bayer Dot Screening

Bayer Screen of Size 16

50

100

16x16 Bayer Dot 150

200
4

6
250
8

10 300

12

14
350

16
2 4 6 8 10 12 14 16 50 100 150 200 250

• Doesn’t look much different than the 8 × 8 case.

• No trade-off between resolution and number of gray lev-
els.
C. A. Bouman: Digital Image Processing - January 8, 2018 14

Example: 128 × 128 Void and Cluster Screen

(1989)

Void and Cluster Screen

50

100

Void and Cluster Dot 150

20
200

40

250
60

80 300

100
350

120

20 40 60 80 100 120 50 100 150 200 250

• Substantially improved quality over Bayer screen.

C. A. Bouman: Digital Image Processing - January 8, 2018 15

Properties of Dispersed Dot Screens

• Eliminate the trade-off between number of gray levels and
resolution.
• Within any region containing K dots, the K thresholds
should be distributed as uniformly as possible.
• Textures used to represent individual gray levels have low
visibility.
• Improved detail rendition.

• Transitions between textures corresponding to different

gray levels may be more visible.
• Not robust to non-ideal output devices
– Requires stable formation of isolated single dots.
C. A. Bouman: Digital Image Processing - January 8, 2018 16

Error Diffusion
• Error Diffusion
– Quantizes each pixel using a neighborhood operation,
rather than a simple pointwise operation.
– Moves through image in raster order, quantizing the
result, and “pushing” the error forward.
– Can produce better quality images than is possible with
screens.
C. A. Bouman: Digital Image Processing - January 8, 2018 17

Filter View of Error Diffusion

f (i, j) + + f˜(i, j) b(i, j)
Quantizer
+ −
+ +

e(i, j)
h(i, j)

• Equations are

255 if f˜(i, j) > T
b(i, j) =
0 otherwise

e(i, j) = f˜(i, j) − b(i, j)

X
f˜(i, j) = f (i, j) + h(k, l)e(i − k, j − l)
k,l∈S

• Parameters
– Threshold is typically T = 127.
– h(k, l) are typically chosen to be positive and sum to 1
C. A. Bouman: Digital Image Processing - January 8, 2018 18

1-D Error Diffusion Example

• f˜(i) ⇒ circles
• b(i) ⇒ boxes

Time = 0 Time = 3

1.0 1.0
0.5 0.5
0 0
i i
−0.5 1 2 3 4 5 −0.5 1 2 3 4 5

Time = 1 Time = 4

1.0 1.0
0.5 0.5
0 0
i i
−0.5 1 2 3 4 5 −0.5 1 2 3 4 5

Time = 2 Time = 5

1.0 1.0
0.5 0.5
0 0
i i
−0.5 1 3 4 5 −0.5 1 2 3 4 5
C. A. Bouman: Digital Image Processing - January 8, 2018 19

Two Views of Error Diffusion

• Two mathematically equivalent views of error diffusion
– Pulling errors forward
– Pushing errors ahead
• Pulling errors forward
– More similar to common view of IIR filter
– Has advantages for analysis
• Pushing errors ahead
– Original view of error diffusion
– Can be more easily extended to important cases when
weights area time/space varying
C. A. Bouman: Digital Image Processing - January 8, 2018 20

ED: Pulling Errors Forward

1. For each pixel in the image (in raster order)
(a) Pull error forward
X
f˜(i, j) = f (i, j) + h(k, l)e(i − k, j − l)
k,l∈S

(b) Compute binary output


255 if f˜(i, j) > T
b(i, j) =
0 otherwise
(c) Compute pixel’s error
e(i, j) = f˜(i, j) − b(i, j)

e(i − 1, j − 1) e(i − 1, j) e(i − 1, j + 1)

e(i, j − 1) f˜(i, j) = f (k, j)+

X
h(k, l)e(i−k, j −l)
k,l

2. Display binary image b(i, j)

C. A. Bouman: Digital Image Processing - January 8, 2018 21

ED: Pushing Errors Ahead

1. Initialize f˜(i, j) ← f (i, j)
2. For each pixel in the image (in raster order)
(a) Compute

255 if f˜(i, j) > T
b(i, j) =
0 otherwise
(b) Diffuse error forward using the following scheme

e = f˜(i, j)−b(i, j) f˜(i, j + 1)

+ = h(0, 1) ∗ e

f˜(i + 1, j − 1) f˜(i + 1, j) f˜(i + 1, j + 1)

+ = h(1, −1) ∗ e + = h(1, 0) ∗ e + = h(1, 1) ∗ e

3. Display binary image b(i, j)

C. A. Bouman: Digital Image Processing - January 8, 2018 22

Commonly Used Error Diffusion Weights

• Floyd and Steinberg (1976)

7/16

3/16 5/16 1/16

• Jarvis, Judice, and Ninke (1976)

7/48 5/48

3/48 5/48 7/48 5/48 3/48

1/48 3/48 5/48 3/48 1/48

C. A. Bouman: Digital Image Processing - January 8, 2018 23

Floyd Steinberg Error Diffusion (1976)

• Process pixels in neighborhoods by “diffusing error” and

quantizing.
Original Image Floyd and Steinberg Error Diffusion

50 50

100 100

150 150

200 200

250 250

300 300

350 350

50 100 150 200 250 50 100 150 200 250

C. A. Bouman: Digital Image Processing - January 8, 2018 24

Quantization Error Modeling for Error

Diffusion

f (i, j) + + f˜(i, j) b(i, j)

Quantizer
+ −
+ +

e(i, j)
h(i, j)

• Quantization error is commonly assumed to be:

– Uniformly distributed on [−0.5, 0.5]
– Uncorrelated in space
– Independent of signal f˜(i, j)
– E [e(i, j)] = 0
δ(k,l)
– E [e(i, j)e(i + k, j + l)] = 12
C. A. Bouman: Digital Image Processing - January 8, 2018 25

Modified Error Diffusion Block Diagram

• The error diffusion block diagram can be rearranged to
facilitate error analysis

f (i, j) + + f˜(i, j) b(i, j)

Quantizer
+ −
+ +

e(i, j)
h(i, j)

+ + f˜(i, j)
f (i, j)
+ +− b(i, j)
+

h(i, j) e(i, j)

e(i, j)

f (i, j) + +
b(i, j)
−
δ(i, j) − h(i, j)

e(i, j)
C. A. Bouman: Digital Image Processing - January 8, 2018 26

Error Diffusion Spectral Analysis

• So we see that
b(i, j) = f (i, j) − (δ(i, j) − h(i, j)) ∗ e(i, j)

rewriting ...
f (i, j) − b(i, j) = (δ(i, j) − h(i, j)) ∗ e(i, j)
| {z } | {z }
high pass filter quantization
error
– Display error is f (i, j) − b(i, j)
– Quantization error is e(i, j)
– Display error is a high pass version of quantization er-
ror
– Human visual system is less sensitive to high spatial
frequencies
C. A. Bouman: Digital Image Processing - January 8, 2018 27

Error Image in Floyd Steinberg Error Diffusion

• Process pixels in neighborhoods by “diffusing error” and
quantizing.
Original Image Quantizer Error Image

50 50

100 100

150 150

200 200

250 250

300 300

350 350

50 100 150 200 250 50 100 150 200 250

C. A. Bouman: Digital Image Processing - January 8, 2018 28

Correlation of Quantization Error and Image

• Quantizer error spectrum is unknown
• Quantizer error model
E(µ, ν) = ρF (µ, ν) + R(µ, ν)
= ρ(Image) + (Residual)
– ρ represents correlation between quantizer error and
image
Weight ρ
1-D 0.0
Floyd and Steinberg 0.55
Jarvis, Judice, and Ninke 0.8

• Using this model, we have

B(µ, ν) = F (µ, ν) − (1 − H(µ, ν)) E(µ, ν)

= [1 − ρ (1 − H(µ, ν))] F (µ, ν) + noise

• This is unsharp masking

C. A. Bouman: Digital Image Processing - January 8, 2018 29

Additional Topics
• Pattern Printing
• Dot Profiles
• Halftone quality metrics
– Radially averaged power spectrum (RAPS)
– Weighted least squares with HVS constrast sensitivity
function
– Blue noise dot patterns
• Error diffusion
– Unsharp masking effects
– Serpentine scan patterns
– Threshold dithering
– TDED
• Least squared halftoning
• Printing and display technologies
– Electrophotographic
– Inkjet