Scribd Logo
Upload

Welcome to Scribd!

  • Dithering With Blue Noise

    Uploaded by

    snackadmiral
    13 views24 pages
    By Robert A Ulichney IEEE January 1988

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Report this Document
    By Robert A Ulichney IEEE January 1988

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Flag for inappropriate content
    Download as pdf
    13 views24 pages

    Dithering With Blue Noise

    Uploaded by

    snackadmiral
    By Robert A Ulichney IEEE January 1988

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Flag for inappropriate content
    Download as pdf
    of 24
    Dithering with Blue Noise ee ROBERT A. ULICHNEY Dipl altoning, sso rtered to as spatial dithering s the imathd of render the son of contcusone pues on tapas at ae capable of only producing binay tur ee ‘emt The concept of hue nahh egueny white ose iSinuodvced and found to have desable preys aering Ets siganthms for tering sth Bie role are develope, Inve on peurbed enor dition The nature of ier fates produces extensively examined inte Requency domain New ‘res Yor analy the fequeney contro pend patera [erbothrecangur tnd heaporal pds developed be ne ‘linens ound tobe Weal cate to recengter gs Seve) Carta selected ay produced examples Se ndosea Grossaty oF SvavoUs 4. Width of the annuli over which the power spectrum estimate B(f) is averaged to form the radially averaged spectrum Pi). Principle radial wavelength in a homoge: neously distributed field of binary pels rep- resenting the constant gray level g. 493, Variance ofan individual output binary pixel resulting. from halftoning homogeneous region of gray level g. ln] Error filter. Governs how past quantization errors are negatively distributed or “di fused" into the yet to be quantized image in the error diffusion algorithm, £ Continuous space frequency vector, f Principal radial frequency in afield of homo- Beneously distributed binary pixels repre- senting gray level g. 4, Radial frequency. The scalar distance in fre- {quency units from zero frequency in a two «dimensional Fourier transform. Gray level. It has a continuous value in the range 0 (white) to 1 (back) inclusive. Manuscript received May 21,197; revised October 2, 1967. This paper Is based on part of these submited in pata flilinent Bite equierent othe dere of Onto iasophy nthe Department of Hectieal ngineeringand Computer scencestihe ‘Massachusetts Insute of Technology, Cambridge, MA n Tune 1966 “The author is wth Digital Equipment Corporation, Hudson, MA 17433008, USA, IEEE Log Number 871844 1m] Quantized discrete space Image. The output from a halftoning process 110; 81 The binary output image resulting from halt: ‘toning an image consisting of a fixed gray level fol = g. Ila) Continuous amplitude, diserete-space im age. Input to a halftoning process, -n_Discrete-space spatial index vector, Nf) Number of discrete frequency samples in an annulus about radial frequency f, Pf) Power spectrum. In this report, only the ower spectra of binary output of aperiodic halftone processes on a single-input gray level are considered i) Power spectrum estimate, Pat) Radially averaged power spectrum. Sample mean of the frequency samples of B(}in the annulus, If| ~ f)| < a/2, about f, %¢,). Sample variance of the frequency samples of i) in the annulus If] ~ f.| < 4/2, about f, S'UNPHE) Anisotropy of Pi 5 Sample period. 14, uy. Frequency baseband replication vectors. U Frequency baseband replication matrix, (uy:u4) vast. Spatial sampling vectors V Spatial sampling matrix, (4: ¥ The nature of various types of noise is often described bya color name. The most well known example is “white noise,” so named because its power spectrum is lat across all frequencies, much like the visible frequencies in white light. "Pink noise” is used to describe low-frequency white noise, the power spectrum of whichis fatout to some finite hhighrequency limit. There is even the curious case of “brown noise,” named for the spectrum associated with Brownian motion {13} lntroduced in this paper is “blue noise,” the high-frequency complement of pink noise, Which is shown to be important in the generation of good quality digital halitones,, Digital Halftoningisthe method of rendering the illusion ooreseisesoreoass0100 © 188 Ae ‘of continuous-tone pictures on displays capable of only ‘producing binary picture elements. is sometimes referred to.as spatial dithering, Several approaches to this process are known: Overviews of existing halftone algorithms have ‘been written by Jarvis etal [17 Stoffel and Moreland (30), ‘Stuck (31, and Ulichney (34). When a rendering system, such a5 that of the offset printing process, cannot ac- ‘commodate small isolated black or white pixels, a “clus. tered-dot” screen is needed. Otherwise “dispersed-dot” methods are preferred, and are the focus of this pape Techniques can also be categorized by computational com. plexity as either point or neighborhood processes. Ordered dither is a point process that produces output by comparing a single continuous-one input value to a deterministic periodic array of thresholds. For images on Square grids, the threshold arrays attributed to Bayer [5] havebeen inwide use. Homogeneous threshold arrays have also been developed for hexagonal and asymmetric grids, ‘with the theoretical background for the study and design of these dither patterns available in [34, ch. 67) ‘Schemes which generate blue noise are neighborhood ‘operations. As in two-dimensional white-noise patterns, bluernoise patterns are aperiodic and radially symmetric. Although white-noise patterns do not suffer from the cor- related periodicity of ordered dither the fact that they pos- sess energy at very low frequencies result in a grainy appearance. Blue-noise patterns enjoy the benefits of ape- Fiodic, uncorrelated. structure without low-frequency graininess. ‘The important algorithm known as Error Diffusion is , NO. 1, JANUARY 1988 Fig. 4. The dependence of piue variance oon gray level e r PtP is perfectly radially symmetric, the measure of anisot- 2 topy sHhyPHf) is merely an estimate of the above ratio. RADIAL FREQUENCY Thetefore, an anisotropy of h or —10 dB should be con- 7 sidered “background noise,” and a reference line at this level will appear in each plot. ‘Also, if anisotropy is low, that is, clase to ~10 dB, indi cating good radial symmetry then Pi is effectively afunc- tion of one independent varablef, instead of two variables The variance of Pf) isthatof (12) divided by Nf) assum: ing that each of the N,(f) samples are independent. This ‘eduction in variance as N(f) increases is indeed observed inthe experimental data to be presented. Ny{f) depends on the width of the annull A. As indicated ‘earlier, in this paper all estimates B() will consist of 256° frequency samples. The sizeof A was chosen sothat exactly ‘one sample along each frequency axis fll into each annu- lus. A plotof Nf for(a)rectangularand (b) hexagonal grids is plotted in Fig 5. i PAL feeere ae 1000 = o : oo ; Sh Spr gee ea ea eee 3 Sera ea erage ee i 9) ualy oftteaummae To what eet ae y sEYPH(F) be meaningful metrics From @) and the fact that roa eneauency K = 10 segments are used in the estimate P(f) ” ae Se nts nnn ies ai ce eto). ee Pi 70 (12) Go Regular hexagonal grid ee ‘The irregularities in the shape of these plots are a con- sequence of rectangular and hexagonal grids not being per fectly radially symmetric. Thenumber of grid points tha all Into a particular annulus essentially increases linearly, as fone would expect, upto the largest annulus that will com: pletely fit within the shape of the baseband: this occurs at fUS~" = } or rectangular grids and 4,/5°' = 13 for hex ‘agonal grids IML Ware Noise Brnweainc In this section, the process of creating a dispersed-dot halftone by the point process of thresholding an input image with uniformly distributed, uncorrelated (white) noise isinvestigated, The quality of output fram this method does not deserve consideration for practical use; the wel: known method of dispersed dot ordered dither performs much better than this one, taking no more computational effort So why shoulda section be devoted to this socalled tech nique of “random dither”? There are two reasons Thefitstishistorical, The dea was the first used to exploit the fact that electronic displays can have independently addressable dots. Goodall [15] in 1951 and Roberts (24) in 1962 demonstrated how contouring due to insufficient gray levels can be corrected by adding noise ofthis type. This 's perhapsthe frst technique that comes to mind to correct the shortcomings of usinga fixed threshold, andinthe early days of digital halftoning it was always referenced for com- parison: in fact the name “ordered dither” was meant to Eontrast random dither Secondly, investigating the output of single gray levels 119; gl dithered in this way, provides a means to check the validity of the newly introduced metrics of radially averaged ower spectrumandanisotropy. Since Jn: gliswhite noise, ithasa known autocorrelation function, namely an impulse atthe oxigin with area o3. So, the power spectrum should be raially symmetric with fixed amplitude o}. Such radial symmetry has been observed optically in Fraunhofer dit {action patterns of randomly distributed apertutes (16 ‘The random numbers used in this section are, strictly pseudo-random. They are produced by means of licative ongruental random number generator(19] available with many programming libraries. This isa very efficient scheme requiring only one multiplication and div sion per random number, andin thecase usedin this study, hhasa repeat period of 2 The effects of dithering with white noise on regular rec: ‘angular and hexagonal grids wll now be considered sep: arately A. Rectangular Grids random dithered gray-scale ramp is shown in Fig. 6.An ‘example ofthe effect of random dither on ascanned image [2tisgivenin Fig. 7. They suer from a grainy appearance. This isthe case at any displayed resolution because of the presence of long wavelengths (low frequencies) tall gray fevels. Both the rectangulatly sampled gray-scale ramp and scanned image contain 300 by 400 pixels Theratially averaged power specteumandanisotropyare ‘essentially the same forall gray levels, and is show in Fig. ‘84a for one particular gray level g nature of these plots validate four things: Fi 6 Rectangular random dither ofa gray-scale ramp, 1) The amplitude of Pf) is correct, that fs, flat, as ‘expected for white noise 2) The values of 6} are as predicted by (3), Fig. 4 i | e Fig.8._Anayss of random ther: Radia spectra in units fo, and anisotropy in stMluncion (dounitsofS~ forthe fxedgraylevelg = sithef = 0.054 () Rectangular ase) Hexagonal case 3) The apparent variance of Pf) decreases with NAF) (Fig, 5} with a minimum at}, (1 ¥/3for the hexagonal case) ‘4 Theanisotropy measureiscortect. Whitenoiseisadi- ally symmetsc so the anisotropy should be at the “back round noise” minimum of ~10 dB. ‘The characteristic features seen here will serve asa eet cerence far the other plots to be studied, B. Hexagonal Grids Hexagonal radially averaged power spectra exhibit the same well-behaved features asin the rectangular case. A random dithered gray-scale ramp and seanned image are shown in Figs, 9and 10. The continuous-tone source images used were the same as those used in the rectangular case, Fig. 9. Hexagonal random dither ofa gray-scale ramp, ‘a Fig. 10, Hexagonal random dither ofa seanned p and displayed withthe same number of samples per unit area, These hexagonally sampled images contain 273 by 40 pixels, with alternate rows offset by one half horizontal period, ‘Again, the radially average power spectrum and anisot ropy are essentially the same fr all gay levels. One rep- resentative sample at g = jis shown in Fig. 8) IV. Princirts Waveucre Consider the problem of rendering a fixed gray level g with binary pixels on regular rectangular or hexagonal grids. For gray levels between g = 0 (white) and g = J, the black pixels are in the minority and itis their spatial distribution ‘onabackground of white that determine the perceived pat terns. The patterns for dark gray levels, between g = 1(black) and g = |, are perceived as white pixels on a black back- round The goal is to distribute the binary pixels as homoge- neously as possible. The pixels, if so arranged, would be Separated by an average distance in two dimensions, This distance is called the Principal Wavelength, and would have the value (na ses] { We Oss} | WWi-g Y

    You might also like