A COMPARATIVE STUDY OF LOSSLESS CODING TECHNIQUES FOR

SCREENED CONTINUOUS-TONE IMAGES

1;  2

Koen Denecker Peter De Neve

ELIS, University of Gent, Sint-Pietersnieuwstraat 41, B-9000 Gent, Belgium

1 2

Email: [email protected] e [email protected] e

ABSTRACT of twotwo-dimensional black-and-white oriented adaptive

co ders: BILEVEL co ding (by Witten et al.) and JBIG

The huge sizes of screened colour-separated photographic

(latest -standard).

images makes very b ene cial for b oth

The compression eciency of these co ders is investigated

storage and transmission. Because of the sp ecial struc-

using six test images, representing halftones at di erent

ture induced by the half-tone dots, the compression results

resolutions, at di erent screening angles and at di erent

obtained on the CCITT test images might not apply to

screening techniques.

high-resolution screened images and the default parameters

of existing compression algorithms may not b e optimal. In

2. DIGITAL COLOUR SCREENING

this pap er we compare the p erformance of di erent classes

of lossless co ders: general-purp ose one-dimensional co ders,

The most common pro cedure in colour screening technology

non-adaptivetwo-dimensional black-and-white co ders and

is as follows [4 ]. Firstly, the CMYK colour image is sepa-

adaptive two-dimensional co ders. Firstly, exp eriments on

rated into its four colour channels (cyan, magenta, yellow

a set of test images screened under di erent conditions

and black). Each colour separation is a greyscale image,



showed that MGBILEVEL and JBIG p erform b est with re-

which is then screened under a di erent angle (e.g., 45 for

 

sp ect to compression eciency; the di erence with the other

the most covering colour black, 15 and 75 for cyan and



co ders is signi cant. Secondly,weinvestigated the in uence

magenta, and 0 for the least visible colour yellow) in order

of the screening metho d (sto chastic or classical screening)

to avoid the app earance of moire patterns. The screening

and screening resolution on the compression ratio for these

step itself converts the original greyscale image into dots

techniques.

placed on a rectangular grid. The sizes of the dots are

determined by the intensity of the greyscale original. This

typ e of screening is called classic screening. Another typ e of

1. INTRODUCTION

screening, stochastic screening, pro duces equally sized dots

at varying distances in order to repro duce the grey tones,

The repro duction of continuous-tone images using the four

giving the dots a random-like app earance.

printing inks (CMYK) consists of splitting the image into

four colour separations and screening each colour separa-

The screening parameters are illustrated in gure 1. It

tion under a di erent angle. New applications, such as shows enlargements of the same detail of the cyan separa-

printing-on-demand, require an almost identical image to tion of the \Musicians" test image screened under di erent

b e screened many times; therefore, storing the unchanged screening metho ds and screening resolutions. The details

part as a screened halftone using lossless compression tech- are part of the test images in our research.

niques instead of screening it again and again, may save

There are two imp ortant parameters for pro ducing digital

printing time as well as storage space. Transmitting such

screens: the screening resolution at which the laser sp ots are

images at the very latest moment also b ene t from com-

pro duced, varying from ab out 300 dpi to ab out 5000 dpi,

pression. Moreover, the sizes of the images (often ab out

and the screen ruling (only in the case of classic screens and

100 MByte) and the time pressure in the application area

expressed in lines p er inch), indicating the size of the grid

(e.g., printing newspap ers) make lossless compression even

and varying from ab out 75 lpi for newspap ers to 300 lpi for

more desirable.

ne art repro ductions. Screening a contone image at high

In this pap er, we have compared the compression e-

resolution increases the image data size drastically, making

ciency of several lossless compression schemes. As a refer-

the need for compression more urgent (e.g., the \Musicians"

ence, wehave exp erimented with four general-purp ose byte-

test CMYK image, 16 Mbyte large, pro duces 200 Mbyte

oriented co ders: GZIP (which is a dictionary-based loss-

after halftoning at a fairly high resolution of 2540 dpi).

less co der), STAT (an optimized PPM-based compressor by

F. Bellard), TIFF PackBits (a simple run-length compres-

3. LOSSLESS COMPRESSION SCHEMES

sor) and TIFF LZW (Lemp el-Ziv-Welch, a dictionary-based

Since each pixel of a screened image is either white or

compression scheme like GZIP). Then, we have used two

black, a screened image is actually is a bilevel image, al-

nonadaptive co ders which accompany the TIFF-standard,

b eit with a very sp ecial structure. The investigated co ders

namely TIFF Group 3 and Group 4 compression (the for-

can b e divided into three classes: one-dimensional general-

mer published fax-standards by CCITT [1 , 2, 3]). Finally,

purp ose co ders (GZIP, STAT, TIFF PackBits and TIFF

wehave also investigated the compression eciency (using

LZW), two-dimensional nonadaptive co ders (TIFF Group 3

standard options and after ne-tuning of the parameters)

and Group 4) and two-dimensional adaptive context-based



Research Assistant with the NFWO, Brussels, Belgium co ders (BILEVEL and JBIG).

(a) (b)

(c) (d)

Figure 1. Magni cation of the same detail of the \Musicians" image: (a) classic screening at 1270 dpi; (b)

sto chastic screening at 1270 dpi; (c) classic screening at 2540 dpi and (d) sto chastic screening at 2540 dpi.

3.1. General-purp ose 1-D byte-oriented co ders 3.2. Nonadaptive 2-D co ders

TIFF Group 3 and TIFF Group 4 are ocially known as

This rst class of co ders do esn't exploit the two-dimensional

CCITT Recommendations T.4 [1 ] and resp. T.6 [2]. These

correlation present in images. Neither do es it take into

compression schemes are not byte-oriented but treat each

account the physical interpretation of the bytes: they are

line of the image as an alternating sequence of black runs

merely interpreted as symb ols of an alphab et, disregarding

and white runs; only the run-lengths are transmitted us-

the numerical value asso ciated with these bytes. On the

ing co de words de ned by the standard, hence the co ders

other hand, they are able to adapt themselves one wayor

are nonadaptive. The co de words are chosen to b e optimal

another to the information stream.

for scanned textual images, but the nonadaptivity could

GZIP (a free and wide-spread compressor by GNU) uses

cause it to p erform badly on images with sp ecial structures.

Lemp el-Ziv co ding (often denoted as LZ77 [5 ]), which is

Group 3 enco ding basically is one-dimensional but provides

a form of sliding-window compression. While running

additional options such like two-dimensional enco din and

through the image le, the window consisting of the last

byte-aligning. Group 4 is a simpli ed or streamlined ver-

n bytes (e.g., n = 1024) is used as a dictionary and each

sion of the Group 3 standard in which only two-dimensional

new symb ol is replaced by an index to this sliding-window

co ding is allowed.

dictionary.

STAT uses a context (i.e., the string consisting of the

3.3. Adaptive 2-D context-based co ders

last 4 symb ols) to feed an arithmetic co der [6 ]. Since it is

These co ders use a context, i.e., an amount of previous pix-

imp ossible to construct complete conditional probabilities

els, to condition the value of each new pixel. Since these

with a context of 4 bytes, only these sp eci c contexts which

co ders will prove to b e the most ecient ones, we will de-

arise more than once are taken into account. This approach

scrib e them in more detail than the other.

is called PPM-based, which stands for Prediction byPartial

3.3.1. JBIG

Matching [7 ].

JBIG (Joint Bilevel Image Group) is the newest CCITT

TIFF Packbits replaces consequent runs of identical sym-

lossless facsimile compression standard develop ed by the

b ols by the numb er of symb ols and the value of the symb ol;

Joint Bilevel Image Group [3]. It is an adaptive co der which

this technique is called run-length co ding. It is a very sim-

supp orts progressive transmission. It uses the Q-co der as a

ple and fast algorithm.

statistical co der but the fact that it's patented by IBM is the

TIFF LZW (Lemp el-Ziv-Welch) [8 ] is a dictionary-based

most imp ortant explanation for the very small p opularity.

technique like GZIP. The technique uses previously co ded

text as a dictionary. A dictionary consisting of strings is

built, in suchaway that each shorter string also exists in Organization of the image data The image is sepa-

the library. The enco ded les consists of indexes to the rated into resolution layers in case of progressive transmis-

longest matching strings. sion, bitplanes in case of greyscale images and horizontal AA ? ?

??

Figure 2. JBIG provides two di erenttyp es of con-

text templates.

Figure 3. Two context templates are provided: the

left one contains 10 pixels and is always used, the

strip es. Co ding will take place p er layer, p er plane and p er

right one contains 22 pixels but is only used when

strip e, called a SDE (Strip e Data Entity).

necessary.

7

Contexts A context is de ned as a numb er of neighb our-

Di erent context templates are chosen for the

ing pixels. 6

west resolution and the other layers. The lowest reso-

lo 5

lution template uses 10 pixels, resulting in 1024 di erent

4

contexts. Two di erent templates are prop osed, as can b e

seen in gure 2; the left one uses more memory but is more 3

t. The current pixel is denoted as \?". Note the ex-

ecien 2

istence of an adaptive pixel \A", which can b e chosen freely

1

in a much larger neighb ourho o d of the current pixel. This

Average compression ratio

be very adequate when compressing p erio dical struc-

can 0

tures. The templates at the higher resolution layers use 4 bil gz g3 g4 pb bie lzw bil+ gz9 stat bie+

pbm

pixels in the lower resolution layer and 6 pixels in the cur-

Codec

rentlayer. Two more bits are added to indicate the spatial

fase, resulting in 4096 contexts.

Figure 4. Average compression ratio for the di er-

ent co decs on the halftone test images (pbm indi-

Resolution reduction A simple way to pro duce a lower-

cates the uncompressed le).

resolution image at a quarter of the original size can be

achieved by subsampling. Since problems with aliasing or

detected run of \1" bits reaches a prede ned length. The

image quality ( ne details get lost) can arise, another lter

multiplication is avoided by rounding o probability esti-

is used, which uses 9 pixels in the currentlayer and 3 pixels

mations, which has a minimal e ect on co ding eciency.

in the lower-resolution layer to pro duce a new pixel in the

lower-resolution layer.

3.3.2. BILEVEL

BILEVEL is a similar but simpler technique [7 ] which

Deterministic prediction A situation can o ccur that

uses twovariable-sized templates as contexts. The smaller

the value of a current pixel can b e predicted in a determin-

context (normally 10 pixels) is always used to condition the

istic way based up on the already deco ded pixels in the cur-

new pixel value, the larger context (normally 22 pixels) is

rentlayer and the values of the pixels in the lower-resolution

only used when it has o ccurred at least twice, see gure 3.

layer. Then, the value of the current pixel is not co ded.

This approach is similar to Prediction byPartial Matching.

The two templates can be changed by the user and opti-

mized for a certain application. The advantage of using big

Typical prediction This typ e of prediction is very useful

contexts is that a b etter prediction can be obtained; the

when large homogeneous regions app ear in images. In the

disadvantage is that it takes much longer to build reliable

lowest-resolution layer, a scanline is said to b e typical when

statistics.

it's identical to the previous scanline. In that situation, a

sp ecial co de word is generated instead of co ding the entire

4. EXPERIMENTAL RESULTS

line and the two corresp onding lines in the higher resolution

layer are not co ded. In the other layers, a pixel is said to

Six test images have b een investigated, representing dif-

be typical when it equals its 8 neighb ours as well as the 4

ferent screening technologies, di erent screening angles

corresp onding higher-resolution pixels.

and di erent screening resolutions. All test images were

screened from the same CMYK continuous-tone original

\Musicians".

Arithmetic co der The JBIG compressor uses IBM's Q-

The compression ratios for the di erent co decs, averaged co der as a statistical co der [9 ]. This co der is a fast approx-

unweightedly over the set of test images, are shown in g- imation of a binary arithmetic co der, i.e., the symb ols can

ure 4. Some co decs provide options by which they can b e only take on the values of MPS (More Probable Symb ol)

optimized for eciency: \gz9" stands for GZIP -9 and is an and LPS (Less Probable Symb ol). The co de word is de ned

optimized but slower version of GZIP, \bil+" is the same as the binary representation of the lower limit of the cur-

as BILEVEL except that the large template has the size rentinterval after the entire sequence has b een co ded. This

of a half diamond and counts 42 pixels instead of 22 pix- lower limit only changes when the MPS is encountered and

els and \bie+" means JBIG with several multi-resolution can b e achieved by adding the interval corresp onding with

decomp osition options altered (a maximal horizontal o set the LPS. Fixed precision arithmetic is used, which means

for the adaptive pixel of 16 pixels; 3 di erential resolution that a renormalisation must b e applied to b oth co de string

layers; maximal 20 lines in each strip e in the lowest resolu- and interval, at co der and deco der, to maintain the present

tion layer). The options were optimized for one image and interval with sucient accuracy. Carry propagation e ects

are not necessarily optimal for the other halftones. are controlled by bit-stung, a zero b eing inserted when a

16

2). The di erence between the two latest fax-standards

TIFF Group 4 and JBIG is signi cant, unlike for scanned

12

textual do cuments. Fine-tuning of the parameters of the

latter two shows to b e useful.

8 Compression eciency using these b est metho ds do esn't

dep end on the screening technique at normal resolutions.

wever, increasing the resolution leads to normal be-

4 Ho

haviour of the compression ratio for classic screens, but not

Compression ratio Compression

hastic screens. Because of the physical limitations

0 for sto c

tent

ro1270 ro2540 mo1270 mo2540 of the ink dot in the latter case, the actual image con

increase when the screening resolution reaches a

Image do esn't

certain level.

Figure 5. In uence of screening metho d and screen-

ACKNOWLEDGEMENTS

ing resolution on compression ratios (only the av-

This work was nancially supp orted by the Belgian Na-

erage of JBIG+ and BILEVEL+ is shown): \ro"

tional Fund for Scienti c Research (NFWO), through a

indicates classic screening, \mo" indicates sto chas-

mandate of Research Assistant and through the pro jects

tic screening, the numb er indicates the resolution.

39.0051.93 and 31.5831.95, and by the Flemish Institute

From this gure, one can see that the nonadaptive tech-

for the Advancement of Scienti c-Technological Research

niques as well as run-length enco ding techniques p erform

in Industry (IWT) through the pro jects Tele-Vision (IWT

worst (average compression ratio less than 3); that the one-

950202) and Samset (IWT 950204).

dimensional, general-purp ose, byte-oriented co ders GZIP

and STAT p erform remarkably well (average compression

REFERENCES

ratio of 4) and that the b est results can b e exp ected from

[1] The International Telegraph and Telephone Consulta-

BILEVEL and JBIG (average compression ratio of 7). A

tive Committee (CCITT), Geneva, Standardization of

larger context size of BILEVEL increases the compression

Group 3 facsimile apparatus for document transmission,

ratio a bit more.

Recommendation T.4, 1985. Volume VI I, Fascicle VI I.3,

Figure 5 shows the compression ratio, averaged over the

Terminal Equipment and Proto cols for Telematic Ser-

two b est compressors, as a function of the screening pa-

vices. Pages 16{31.

rameters. Combinations of di erent screening techniques

[2] The International Telegraph and Telephone Consulta-

(classic and sto castic screening, indicated by\round dots"

tive Committee (CCITT), Geneva, Facsimile Coding

and \monet dots" resp ectively) and di erent screening res-

Schemes and Coding Control Functions for Group 4

olutions (1270 and 2540 dpi) are compared. The size of

Facsimile Apparatus, Recommendation T.6, 1985. Vol-

the uncompressed le is equal for b oth techniques (approx-

ume VI I, Fascicle VI I.3, Terminal Equipment and Pro-

imately 13 Mbyte at 1270 dpi and 50 Mbyte at 2540 dpi).

to cols for Telematic Services. Pages 40{48.

From this gure, one can see that when doubling the reso-

lution in the case of classic screening, the compression ratio

[3] The International Telegraph and Telephone Consulta-

doubles while the umcompressed le increases by a factor

tive Committee (CCITT), Progressive Bi-Level Image

of four, which is considered to b e normal. But, in the case

Compression. Recommendation T.82, 1993.

of sto chastic screening, the compression ratio is almost four

[4] C. Eliezer, \Color screening technology: A tutorial on

times as big, which means that the size of the compressed

the basic issues," The Seybold Report on Desktop Pub-

le almost do esn't increase. This leads to the interpretation

lishing,vol. 6, Octob er 1991.

that the actual image content nearly do esn't increase.

[5] J. Ziv and A. Lemp el, \A universal algorithm for se-

Another conclusion one can draw from this gure is that

quential ," IEEE Transactions on In-

the compression ratios at normal resolution (1270 dpi) do

formation Theory,vol. 23, pp. 337{343, May 1977.

not dep end on the screening technology used, while at

higher resolutions, classic screening adds more noise/detail

[6] F. Bellard, \Compression statistique a contexte ni.,"

to the screened image than sto chastic screening do es. This

June 1995.

is induced byphysical limitations of the ink dot size: though

http://www.p olytechnique.fr/~b ellard/.

the laser sp ot resolution gets doubled, the printing dot size

[7] J. G. Cleary and I. H. Witten, \Data compression us-

remains the same and only the lo cation of the dot can vary.

ing adaptive co ding and partial string matching," IEEE

Transactions on Communications,vol. 32, pp. 396{402,

5. CONCLUSION

April 1984.

In this pap er wehave compared the compression eciency

[8] T. A. Welch, \A technique for high p erformance data

of several existing lossless co ders on screened continuous-

compression," IEEE Computer,vol. 17, pp. 8{19, June

tone images. Wehave shortly explained the most imp ortant

1984.

screening technologies and the investigated compressors.

[9] W. B. Pennebaker, J. L. Mitchell, G. G. Langdon, and

The compressors can b e divided into three classes: adap-

R. B. Arps, \An overview of the basic principles of the

tive 1-D general-purp ose co ders, nonadaptive 2-D co ders

Q-co der adaptive binary arithmetic co der," IBM J. Res.

and adaptive 2-D co ders. Exp eriments with halftones, pro-

Devel.,vol. 32, pp. 717{726, 1988.

duced from the same continuous-tone original under di er-

ent screening conditions, have shown that the adaptive 2-D

co ders p erform b est (average compression ratio of ab out 6),

followed by the adaptive 1-D general-purp ose co ders (av-

erage compression of ab out 4) and the nonadaptive 2-D

co ders p erform worst (average compression ratio of ab out