LOSSY COMPRESSION OF STOCHASTIC HALFTONES WITH JBIG2
Magesh Val liappan and Brian L. Evans Dave A. D. Tompkins and Faouzi Kossentini
Dept. of Electrical and Computer Eng. Dept. of Electrical and Computer Eng.
The UniversityofTexas at Austin University of British Columbia
Austin, TX 78712-1084 USA Vancouver, B.C. V6T 1Z4 Canada
fmagesh,b [email protected] fdavet,[email protected] c.ca
ibility for enco der designs. An enco der could segment bi- ABSTRACT
level images into text, halftone, and generic regions, and
The JBIG2 standard supp orts lossless and lossy co ding mo d-
enco de each region separately [3, 4]. For halftone regions,
els for text, halftone, and generic regions in bi-level im-
JBIG2 supp orts very high lossy compression rates. This
ages. For the JBIG2 lossy halftone compression mo de,
pap er develops a new high-quali ty lossy JBIG2 enco ding
halftones are descreened b efore enco ding. Previous JBIG2
metho d tailored for sto chastic halftones. We develop visual
descreening implementations pro duce high-quali ty images
quality measures, whichwe use for rate-distortion tradeo s.
for clustered dot halftones at high compression rates but
Weachieve an additional 35 compression over the b est re-
signi cantly degrade the image quality for sto chastic half-
p orted metho d [5]. We implement the metho d in the JBIG2
tones, even at muchlower rates. In this pap er, we develop
co dec available at http://spmg.ece.ubc.ca/jbig2.
1 a exible, computational ly ecient, JBIG2-compliant
metho d for compressing sto chastic halftones that reduces
2. BACKGROUND
noise, artifacts, and blurring; 2 quality measures for lin-
ear and nonlinear distortion in compressed halftones; and
JBIG2 uses a form of vector quantization for compressing
3 rate-distortion tradeo s for the enco der parameters.
halftones. First, the enco der descreens inverse halftone
the bi-level image into a grayscale image with reduced spa-
1. INTRODUCTION
tial resolution. Gray levels serve as indices into a halftone
pattern dictionary, whichischosen by the enco der. Then,
Digital halftoning converts a continuous-tone image into a
the enco der enco des the halftone pattern dictionary and
bi-level image halftone for printing and display on binary
grayscale image bitplanes losslessly using the generic mo de
devices. Bi-level images consist of a single rectangular bit
of op eration, and cho oses the orientation for the halftoning
plane. Pixels are either assigned black or white to create
grid. The deco der rst deco des the bitplanes to construct
an illusio n of continuous shades of gray.
the grayscale image and then constructs the bi-level im-
Halftoning by ordered dithering thresholds a grayscale
age by placing the dictionary patterns corresp onding to the
image by using a p erio dic mask of threshold values. In
grayscale values at their appropriate p ositions and orienta-
clustered dot ordered dithering, black dots are clustered
tions. Thus, the spatial resolution and numb er of gray levels
in large blobs. Since clustered dot halftones are resistant
directly a ect the quality and compression ratio. Perceptu-
to ink spread, they are the most common among printed
ally lossless compression is achieved by preserving the lo cal
halftones. Sto chastic halftoning varies the thresholding ac-
average gray level but not the bi-level image itself.
cording to lo cal statistics in the image in order to shap e
One descreening metho d [5] maps each non-overlappin g
the quantization noise into the high frequencies where the
M M window of halftone pixels to one grayscale pixel
human visual system is less sensitive. Sto chastic halfton-
value that is equal to the numb er of black pixels in the
ing requires more computation and generally yields b etter
halftone window. For non-angled grids, this metho d yields
2
visual quality than ordered dithering [1].
M + 1 gray levels. For angled grids, a mask is used with
Current fax machines only supp ort one lossless com-
the window so that adjacent patterns do not overlap, and
2
pression mo de that has b een optimized for textual data [2]
the numberofgray levels is area of the mask +1 M +1.
and causes data expansion when applied to halftones. The
This metho d is conceptually and computationall y simple,
Joint Bi-Level Exp erts Group JBIG is a sub committee of
and works well for ordered dithered halftones. When this
b oth the ISO/IEC and the ITU-T that is developing a sec-
metho d is applied to sto chastic halftones, the grayscale im-
ond international standard for bi-level image compression
age su ers from noise, blur, and artifacts, which degrade
for use in printers, fax machines, scanners, and do cument
the reconstructed halftone quality[5].
storage and archiving. JBIG2 adds lossy compression and
For sto chastic halftones, inverse halftoning metho ds in
supp orts several di erent co ding mo dels for text, halftone
descending order of quality are set theoretic [10], nonlinear
and generic regions. JBIG2 should b e nalized in Fall 1999.
denoising [7], adaptive smo othing [6], overcomplete wavelet
JBIG2 sp eci es the bit stream syntax, which places
expansion [8], wavelet denoising [9], pro jection onto convex
strict requirements on deco der designs but leaves much ex-
sets [11], and vector quantization [12]. The inverse halftones
have the same spatial resolution as the halftone but with
M. Valliappan and B. Evans were supp orted by a US Na-
6{8 bits of precision p er pixel. For higher compression,
tional Science Foundation CAREER Award under grant MIP-
a JBIG2 enco der would create a grayscale image at lower
9701707. D. Tompkins and F. Kossentini were supp orted by the
spatial resolution and fewer gray levels. Natural Sciences and Engineering Research Council of Canada.
3. PROPOSED ENCODING METHOD 4. QUALITY METRICS
We develop quality metrics to evaluate the p erformance
The prop osed enco der in Fig. 1 takes a sto chastic halftone
of the prop osed enco ding metho d. The prop osed enco d-
as input and generates a JBIG2-compliant bitstream as out-
ing metho d attempts to preserve the useful information
put. The free parameters are grid size and orientation, num-
present in the sto chastic halftone while discarding as much
b er of quantization levels, and sharp ening control.
noise and distortion as p ossible. So, we compare the input
The pre lter should suppress high-frequency noise, spu-
halftone with the original grayscale image.
rious tones, and Nyquist frequencies in sto chastic halftones,
Signal to noise ratios SNRs assume that the only source
and it should have a at passband resp onse to minimize
of degradation in a pro cessed signal is additive noise. Com-
distortion [13]. To meet these criteria while maintaining
pression and halftoning also intro duce linear and nonlinear
computational simplicity,we design a symmetric 3 3 -
distortion. Because the human visual system resp onds in-
nite impulse resp onse lter with p ower-of-two co ecients
"
1 2 1 dep endently to linear distortion and noise, we develop a
[13]: 2 4 2 . The pre lter pro duces a grayscale image quality measure for each e ect [13 , 18]. We quantify the
1 2 1 linear distortion by constructing a minimum mean squared
at the same spatial resolution as the halftone.
error Weiner lter so that the residual image is uncorre-
The decimator consists of a lowpass anti-aliasing lter
lated with the input image. The residual image represents
followed bydownsampling by M in each dimension. The
the nonlinear distortion plus additive indep endent noise.
lter sums the pixel intensities within the pattern mask.
To quantify the e ect of nonlinear distortion and noise
For unangled grids, this corresp onds to a cuto frequency
on quality,we sp ectrally weight the residual by a contrast
of in each dimension. This pro cess is very similar to
sensitivity function CSF. A CSF is a linear approximation
M
[5] except that the mask and the window are applied to
of the human visual system resp onse to a sine waveofa
the ltered grayscale values instead of the bi-level image
single frequency [13, 18]. A lowpass CSF assumes that the
and that the numb er of resulting gray levels is signi cantl y
observer do es not fo cus on one p oint in the image but freely
higher. In our implementation , we combine the pre lter
moves the eyes around the image. We form a weighted SNR
and the anti-aliasin g lter. Since the input would b e a bi-
P P
2
jX u; v C u; v j
level image, we replace multiplicati ons with additions. For
u v
P P
WSNR = 10 log
10
2
a33 smo othing lter and a window size of M M ,we
jD u; v C u; v j
u v
2
need at most M +2 additions for each grayscale value.
2
The quantizer uses N gray levels N M + 1. Con-
where C u; v isalowpass CSF, and X u; v and D u; v
ventional quantization with 32 or fewer levels M 5
are the Fourier transforms of windowed original and residual
would create false contours [14]. Toavoid contouring, we
images, resp ectively. A higher WSNR means higher quality.
dither the input image using a multi-level version [15 ] of
To prevent WSNRs to b e biased by large DC comp onents,
mo di ed error di usion [16]. Error di usion, whichisa
we initiall y remove the DC comp onent of the images.
typ e of 2-D sigma-delta mo dulation, shap es the quantiza-
To quantify the e ect of linear distortion, we compute
tion noise into the higher frequencies. Mo di ed error dif-
a Linear Distortion Measure
fusion controls the amount of linear distortion by means of
P P
j1 H u; v jjX u; v C u; v j
a sharp ening control parameter L, which requires an extra u v
P P
LDM =
jX u; v C u; v j
addition and multiplicati on p er input pixel. Wecho ose L
u v
to comp ensate for the blurring in the previous stages. To
where H u; v is the Fourier transform of the Weiner lter
achieve higher compression, we could further quantize the
that mo dels the pro cess. H u; v islowpass, so 1 H u; v
gray levels with a slight loss in quality. In our implementa-
is highpass. Because of the weighting by X u; v , the LDM
tion, we use a four-tap Floyd-Steinb erg error di usion lter
only measures those frequencies which are present in the
which has dyadic co ecients. It is the smallest error di u-
original image. A higher LDM means lower quality.
sion lter known to pro duce high-quality halftones [13, 17 ].
The size, shap e, and orientation of the patterns directly
5. RESULTS
a ect the reconstructed halftone quality. Halftone patterns
o
oriented at 45 can yield p erceptually b etter results [1]. The
We compress the 512 512 Floyd-Steinb erg error di used
p ossible patterns dep end on the rendering device. We use
halftone of the grayscale image barbara shown in Fig. 2.
metho ds similar to [5] to generate halftone patterns for b oth
Figs. 3{9 show the e ect of the enco ding parameters. Table
angled and non-angled clustered dot masks.
1 gives the compression rates and distortion. The distortion
measures assume a 600 dpi rendering device and a 40 cm
viewing distance. For lossy compression, the enco der uses a y clustered dot halftones [5]. Lossless pattern dictionary generated b PrefilterDecimator Quantizer
Encoder
As shown in Table 1, the existing Group 4 MMR fax
standard expands the original image by 148 whereas the
arithmetic JBIG2 generic MQ lossless co der gives a com-
If the image is enco ded without Symbol pression ratio of 1.73.
Dictionary
a pre lter, then the noise intro duced by sto chastic half-
toning is visible Fig. 3. By pre ltering, we reduce the
Figure 1: Prop osed JBIG2 enco ding metho d for compress-
high-frequency noise and some of the detail Fig. 4. We
ing halftones. Free parameters are pattern grid size M ,
comp ensate for the frequency distortion by adjusting the
numb er of quantization levels N , and sharp ening control L.
sharp ening parameter Fig. 5. The sharp ening parameter
The pre lter would b e applied only for sto chastic halftones.
improves visual quality but decreases the compression ratio
Figure 2: Original halftone of the barbara image. Figure 3: Reconstructed halftone without pre ltering using
a44 grid M = 4 and N = 17 gray levels. Same as [5].
Fig. 4{6. Fig. 6 depicts oversharp ening contrast distor-
tion.
Using a larger grid size and an angled screen can achieve
b etter quality for the same bit rate Figs. 5 and 7. In-
creasing the grid size increases compression Figs. 7 and 8.
Quantizing the descreened image to fewer gray levels further
increases compression Fig. 9. Rate-distortion curves Fig.
10 represent the variation of distortion and compressed im-
age size with resp ect to the grid size and sharp ening control
parameter. For linear distortion, an optimal value of the
sharp ening control parameter exists. Beyond the optimal
value, the reconstructed halftones are oversharp ened and
distortion increases, which increases b oth rate and distor-
tion. WSNR do es not change signi cantly with the sharp-
ening control parameter but changes dramatically with the
grid size. The sharpness control parameter can b e used to
trade o compressed image size for improved quality.
Enco der Parameters Distortion Size
Image L M N LDM WSNR Bytes
Orig. { { { { { { 32768
Figure 4: Reconstructed halftone with a pre lter using a
MMR { { { { { { 81420
4 4 grid M = 4 with all 17 gray levels N = 17.
MQ { { { { { { 18907
o
Fig. 3 { 4 17 0 0.163 15.4 dB 5374
o
Fig. 4 { 4 17 0 0.181 16.5 dB 4356
6. CONCLUSION
o
Fig. 5 0.5 4 17 0 0.091 16.0 dB 5152
o
Fig. 6 1.5 4 17 0 0.292 14.8 dB 6352
Wehave develop ed a JBIG2-compliant metho d for enco d-
o
Fig. 7 0.5 6 19 45 0.116 18.7 dB 4961
ing sto chastic halftones. To descreen the halftone, we pre-
o
Fig. 8 0.5 8 33 45 0.155 15.7 dB 3983
lter, decimate, and quantize. Pre ltering reduces high-
o
Fig. 9 0.5 8 16 45 0.158 14.0 dB 3318
frequency noise, spurious tones, and Nyquist frequencies.
Decimation reduces spatial resolution. Quantization uses
Table 1: Visual quality and size of the JBIG2 bitstream for
mo di ed Floyd-Steinb erg error di usion to shap e the quan-
the barbara image for the enco der parameters: sharp ening
tization error into the higher frequencies and control the
control L, M M blo ck size, N quantization levels, and
sharpness. We analyze the rate-distortion tradeo s for the
orientation . Section 4 de nes LDM and WSNR. MMR
enco der parameters: orientation of the halftone grid; deci-
and MQ are lossless JBIG2 mo des. MMR is same as the
mation factor which also sp eci es the halftone grid; number
current fax standard. MQ uses arithmetic co ding.
of quantization levels; and sharp ening control. To measure
Figure 5: Reconstructed halftone with pre ltering and Figure 7: Reconstructed halftone with pre ltering and
sharp ening L =0:5 using a 4 4 grid M = 4 with sharp ening L =0:5 using a 6 6 grid M = 6 angled
o
all 17 gray levels N = 17. at 45 with all 19 gray levels N = 19.
Figure 6: Reconstructed halftone with pre ltering and Figure 8: Reconstructed halftone with pre ltering and
sharp ening L =1:5 using a 4 4 grid M = 4 with sharp ening L =0:5 using an 8 8 grid M = 8 angled
o
all 17 gray levels N = 17. at 45 with all 33 graylevels N = 33.
distortion, we develop a quality measure for linear distor-
tion and one for nonlinear distortion and additive noise.
For the same level of distortion, our metho d can achievean
additional 35 compression over the metho d in [5], as seen
by comparing the results Figs. 3 and 8 in Table 1.
7. REFERENCES
[1] R. Ulichney, Digital Halftoning. MIT Press, 1987.
[2] Fascimile Coding Schemes,vol. VI I. Geneva, Switzerland:
CCITT Rec. T.4 T.6, 1984.
[3] K. Denecker, S. Van Assche, and I. Lemahieu, \A fast au-
to correlation based context template selection scheme for
lossless compression of halftone images," Proc. SPIE Color
Imaging Conf.,vol. 3300, pp. pp. 262{272, Jan. 1998.
[4] D. A. D. Tompkins and F. Kossentini, \A fast segmentation
algorithm for bi-level image compression using JBIG2," in
Proc. IEEE Int. Conf. Image Proc., Oct. 1999.
[5] B. Martins and S. Forchhammer, \Halftone co ding with
Figure 9: Reconstructed halftone with pre ltering, sharp- JBIG2," IEEE Trans. Image Proc., to app ear.
ening L=0.5 and quantized to N =16gray levels using
[6] T. D. Kite, N. Damera-Venkata,B.L.Evans, and A. C.
o
an 8 8 grid M = 8 angled at 45 .
Bovik, \A high quality, fast inverse halftoning algorithm for
error di used halftones," in Proc. IEEE Int. Conf. Image
Proc.,vol. 2, pp. 64{8, Oct. 1998. 0.4
8
enkata, T. D. Kite, M. Venkataraman, and [7] N. Damera-V Sharpness Parameter
0.35 L = 0.0
B. L. Evans, \Fast blind inverse halftoning," in Proc. IEEE
7 L = 1.0
Int. Conf. Image Proc.,vol. 2, pp. 71{4, Oct. 1998.
0.3 6
[8] J. Luo, R. de Queiroz, and Z. Fan, \A robust technique for
avelet transform," IEEE
image descreening based on the w 0.25 5
Trans. Signal Proc.,vol. 46, pp. 1179{94, Apr. 1998.
0.2
LDM
[9] Z. Xiong, M. Orchard, and K. Ramachandran, \Inverse half-
toning using wavelets," in Proc. IEEE Int. Conf. Image
0.15
Proc.,vol. 1, pp. 569{72, Sept. 1996.
verse halftoning," Proc. IEEE [10] N. T. Thao, \Set theoretic in 0.1
4
Int. Conf. Image Proc.,vol. 1, pp. 783{6, Oct. 1997.
0.05
3
[11] S. Hein and A. Zakhor, \Halftone to continuous-tone conver-
2
n co ded images," IEEE Trans. Image sion of error-di usio 0
2000 3000 4000 5000 6000 7000 8000 9000 10000 11000
oc.,vol. 4, pp. 208{16, Feb. 1995.
Pr Compressed Image Size (bytes)
a Linear distortion measure
[12] J. Z. C. Lai and J. Y. Yen, \Inverse error-di usion using
ector quantization," IEEE Trans. Image Proc.,
classi ed v 0 ol. 7, pp. 1753{8, Dec. 1998. v 8 Sharpness Parameter L = 0.0
7
Design and Quality Assessment of Forward and [13] T. D. Kite, L = 1.0
5
Inverse Error Di usion Halftoning Algorithms. PhD thesis,
ersityofTexas, Austin, Texas, Aug. 1998.
The Univ 6
t, The Reproduction of Colour.Fountain [14] R. W. G. Hun 10
Press, 1995. 5
WSNR
[15] T. Morita and H. Ochi, \Progressive transmission of contin-
15
uous tone images using multi-level error di usion metho d,"
rans. Comm.,vol. 82-B, pp. 103{11, Jan. 1999.
IEICE T 4
hbach and K. Knox, \Error-di usion algorithm with [16] R. Esc 20
3
edge enhancement," J. Opt. Soc. Am. A,vol. 8, pp. 1844{
2 50, Dec. 1991.
25
yd and L. Steinb erg, \An adaptive algorithm for spa-
[17] R. Flo 2000 3000 4000 5000 6000 7000 8000 9000 10000 11000
yscale," Proc. Soc. Image Display,vol. 17, no. 2,
tial gra Compressed Image Size (bytes)
pp. 75{7, 1976.
b Weighted signal-to-noise ratio
[18] N. Damera-Venkata, T. D. Kite, W. S. Geisler, B. L. Evans,
Figure 10: Rate-distortion curves for the barbara image.
and A. C. Bovik, \Image quality assessment based on a
Each curve is plotted byvarying the sharpness parameter
degradation mo del," IEEE Trans. Image Proc., to app ear.
L 2 [0; 1] for a sp eci c value of M 2f2;3;4;5;6;7;8g.