Showing papers presented at "Color Imaging Conference in 1993"

Characterization of Scanner Sensitivity.

Abstract: Color scanners are becoming quite popular as input devices for desktop publishing. In many applications it is desirable to obtain calibrated color from them. In order to properly calibrate the device for a variety of illuminants, it is necessary to estimate the spectral sensitivity of the scanner. This paper describes a set theoretic approach to this problem. This method is shown to have increased accuracy compared to present methods.

Comparing Appearance Models Using Pictorial Images.

Abstract: Eight different color appearance models were tested using pictorial images. A psychophysical paired comparison experiment was performed where 30 color-normal observers judged reference and test images via successive-Ganzfeld haploscopic viewing such that each eye maintained constant chromatic adaptation and inter-ocular interactions were minimized. It was found that models based on von Kries had best performance, specifically CIELAB, HUNT, RLAB, and von Kries.

Rethinking the White Point.

Abstract: In this talk we will discuss calibration transforms that map the XYZ values generated by the same surface under different illuminants. We use the phrase calibration transforms to distinguish between analyses based on the physical properties of surfaces and illuminants, and to distinguish them from appearance transforms based on measurements of color appearance. Calibration transforms describe how the XYZ coordinates measured for a surface change with illumination. Appearance transforms describe how the XYZ coordinates of a particular appearance change with illumination. The change in XYZ values due to calibration and appearance transforms do not generally coincide. Calibration and appearance transforms serve different and useful functions in color management systems. Calibration transforms can be used to correct the device-independent color descriptors of surfaces that have been calibrated for one illuminant but will be rendered under a different illuminant. These calculations play an important role in device calibration. Appearance transforms share a similar computational structure, but they have a different goal and are not XYZ matches. One step in performing automated calibration transforms is to estimate the illuminant spectral power distribution (SPD). Performing the transform is much simpler if one can estimate the illuminant SPD from a three-sensor device, rather than using a spectroradiometer. We first analyze an illuminant estimation method based on using linear models of the illuminant SPD. This solution is taken from the color constancy literature and assumes very little or no information about the objects in the image. In many practical applications the need for accurate calibration transforms outweighs the advantages of algorithms based on little or no information about the surfaces and illuminants in a scene. In this paper we describe how a measurement of a single calibration with known surface target, such as the Macbeth Color-Checker, improves our ability to estimate the illuminant SPD. This method may be useful in practical applications where a single calibration measurement is permitted.

Color Space Selection for JPEG Image Compression.

Abstract: The Joint Photographic Experts Group's image compression algorithm has been shown to provide a very efficient and powerful method of compressing images. However, there is little substantive information about which color space should be utilized when implementing the JPEG algorithm. Currently, the JPEG algorithm is set up for use with any three-component color space. The objective of this research is to determine whether or not the color space selected will significantly improve the image compression. The RGB, XYZ, YIQ, CIELAB, CIELUV, and CIELAB LCh color spaces were examined and compared. Both numerical measures and psycho-physical techniques were used to assess the results. The final resuLts indicate that the device space, RGB, is the worst color space to compress images. In comparison, the nonlinear transforms of the device space, CIELAB and CIELUV, are the best color spaces to compress images. The XYZ, YIQ, and CIELAB LCh color spaces resulted in intermediate levels of compression.

Color Reproduction and Color Vision Modeling.

Abstract: The principles of spectral color reproduction, as exemplified by the Lippmann method, and of trichromatic color reproduction, as used by presentday systems, are reviewed. Colorimetry is also based on trichromatic principles, and provides the basis for the quantitative evaluation of color reproduction in television, photography, and printing. Being metameric, these reproductions are affected by changes in illuminant and observer. They are also sensitive to changes in their viewing conditions; these changes can be represented by color vision models, one of which is briefly described. Spectral Color Reproduction Colors are defined physically by their spectral power distributions. To reproduce colors with the same spectral power distributions as in the original scene is difficult. The most elegant method so far devised for doing this is the interference method invented by Lippmann. However, its low sensitivity and general inconvenience have made it obsolete for many years. Trichromatic Color Reproduction Because the eye does not respond to each wavelength of the spectrum, but only to the light in three broad bands, it is possible to reproduce colors trichromatically, by displaying them as mixtures of red, green, and blue light. Colorimetry Colorimetry also depends on the trichromatic principle. Colors of each wavelength of the spectrum were matched with mixtures of red, green, and blue light by groups of observers, and the results averaged to represent the response of Standard Colorimetric Observers. These data have led to the now universally-used CIE system of colorimetry, with its X,Y,Z tristimulus values, the x,y and u ́,v ́, chromaticity diagrams, and the CIELUV and CIELAB color spaces. Television If a television camera has the same spectral sensitivity functions as those of the three different types of cone in the eye, correct color reproduction does not occur because each display color stimulates more than one of the three different cone types. This is true whether the display device uses a cathode-ray tube, a triple projector, or a liquid crystal array. These unwanted stimulations result in the colors in the reproduction having reduced saturation and distorted hues. These errors can be corrected by passing the signals from the camera through a matrixing circuit which removes the effects of the unwanted stimulations; but this can only succeed within the gamut of reproducible colors. Colors outside this gamut require negative signals which cannot be displayed. The gamut limitation cannot be avoided, because it results from the overlapping nature of the cone spectral sensitivities. Photography In photography, the display normally consists of cyan, magenta, and yellow dye images, and the bands of wavelengths controlled by them also result in unwanted stimulations. But, unlike television, photography can provide no convenient way of introducing matrixing. What is usually done is to increase the wavelength separation of the three spectral sensitivities of the film as compared to those of the eye, and this improves the fidelity of most colors, although it introduces marked hue errors for a few colors such as certain types of blue flower. The dyes used in photography also absorb in parts of the spectrum where they should not do so, and these unwanted absorptions result in colors being too dark; however, the use of colored couplers and interimage effects are very effective in making the necessary corrections. Half-tone and Digital Printing In conventional printing, the modulation of the light in the display by the cyan, magenta, and yellow inks is achieved by varying the size of dots of ink that are small enough not to be resolved by the eye at normal viewing distances. In digital printing, it is necessary to construct these half-tone dots from patterns of micro-dots; to avoid spurious contouring effects, about ten times as many micro-dots are required per unit distance as compared to the number of half-tone dots. For high quality work this requires microdots of about 1/100 mm in diameter. However, by distributing the micro-dots in dithered patterns instead of in clusters, or by using error diffusion techniques (distribution over neighbouring pixels of the errors caused by using too few micro-dots) high quality results can be obtained with larger micro-dots. Metamerism Trichromatic reproductions inevitably result in almost all colors being reproduced with spectral power distributions that are markedly different from those in the original scene. This metamerism means that even if the reproduction matches the original for one illuminant and observer, it will not necessarily do so for others. Non-self-luminous displays, such as liquid crystal television devices, photographic transparencies and reflection prints, printed papers, and desktop publications, are therefore sensitive to changes in the spectral power distribution of their illuminant; and all

Color Sensations in Complex Images.

Abstract: Colorimetric measurements are equally influenced by the reflectance spectrum of the object and the illumination spectrum of the light. The 1931 CIE colorimetric measurements are made one pixel at a time; they integrate the radiances at each wavelength with three colormatching functions so as to generate three Tristimulus Values for one pixel. No information from other pixels in the field of view is used in this calculation. Our everyday experience is that color appearance of objects remain the same, regardless of substantial changes in the spectrum of the illuminant. In other words, everyday experience tells us that an object’s reflectance spectrum controls appearance, while its illumination spectrum has little influence. This paper will review the history of different hypotheses explaining human color constancy and describe techniques for measuring color appearances. It will review important experiments that measure color sensations and new techniques using the introduction of a new patch in a display that destroys color matches. Human color vision is a field phenomenon. Humans calculate color sensations by comparing pixels across the entire field of view. Global changes in reflectance or illumination cause small changes in appearance: Local changes in reflectance or illumination cause large changes in sensation. The spatial interaction of all pixels in the field of view controls human color appearance.

Modelling Reflectance by Logarithmic Basis Functions.

Abstract: Low-dimensional linear models of spectral reflectance have proved very useful (eg.1,2,3,4) because they allow approximations of the reflectance spectrum to be represented by only a few parameters. While 6-8 such parameters may suffice5 for accurate approximations, in many color applications often only 3 or fewer can actually be used because color imagery provides only 3 knowns to a set of equations in which the linear-model parameters are the unknowns. This paper investigates the extent to which non-linear models might improve the accuracy in approximating reflectance given only a 3-band measurement.Rodriguez and Stockham6 describe a non-linear meth-od which extracts the transmittance spectrum of transparency film based on a knowledge of the spectral transmittance functions of the 3 underlying film dyes. Their algorithm calculates the amount of each dye from the response of a 3-band scanner. Since the film obeys Beer’s law, the dye amounts and scanner responses are related non-linearly. Their iterative algorithm estimates the dye amounts based on the scanner responses, constructs the transmittance spectrum corresponding to those dye amounts and then generates a set of predicted scanner responses which are then compared with the actual scanner responses in order to adjust the dye estimates. Their method works well for film; will it also work well for modelling spectral reflectances? This paper presents results showing that in fact low-dimensional, non-linear models of reflectance approximate spectral reflectance somewhat better than linear models of the same dimension. Nonlinear models, however, demand significantly more computation time.